Accounting, Organizations and Society 68-69 (2018) 88e108

Contents lists available at ScienceDirect

Accounting, Organizations and Society

journal homepage: www.elsevier.com/locate/aos

A new era of voluntary disclosure? Empirical evidence on how

employee postings on social media relate to future corporate
disclosures
Jeffrey Hales a, *, James R. Moon Jr a, Laura A. Swenson b
a
Scheller College of Business, Georgia Institute of Technology, Atlanta, GA 30308, United States
b
Sheldon B. Lubar School of Business, University of Wisconsin e Milwaukee, Milwaukee, WI 53201, United States

a r t i c l e i n f o a b s t r a c t

Article history: With the advent of social media, individual public opinions about firms can be more easily accessed and
Received 31 March 2017 aggregated, and recent research suggests that various platforms, such as Twitter, Seeking Alpha, and
Received in revised form Estimize, provide information relevant in predicting future corporate disclosures. Rather than focusing
11 March 2018
on the general public's opinion, we examine a public platform designed to convey insider information -
Accepted 23 April 2018
Glassdoor.com, where employees voluntarily share their opinions on a number of issues, including the
company's near-term business outlook. Using a sample of approximately 150,000 employee reviews, we
extract both employees' explicit assessments of outlook and a measure of their latent outlook derived
from factor analysis. We then examine whether the opinions employees share on social media relate to
future corporate disclosures. In particular, we find evidence that employee opinions are useful in pre-
dicting growth in key income statement information, transitory reporting items (e.g., restructuring
charges), earnings surprises, and management forecast news. While voluntary disclosures about firm
performance have traditionally come from executives, our evidence suggests that rank-and-file em-
ployees are chipping away at upper-level management's exclusive control over that channel.
© 2018 Elsevier Ltd. All rights reserved.

1. Introduction firms to cultivate and disseminate much of the knowledge held by

Voluntary disclosures represent one of the most widely studied
topics in accounting research and, traditionally, this research has
modeled disclosure decisions at the firm level, where a single
manager considers the costs and benefits of disclosing their private
information to the market place.1 This approach is both intuitive
and practical because upper-level management has largely
controlled firm-level communications, such as press releases,
earnings warnings, and other such announcements. Yet employees
throughout an organization, especially when considered collec-
tively, likely hold a wealth of information about the state of their
company (e.g., Babenko & Sen, 2015; Huddart & Lang, 2003), but
there have historically been few channels through which that in-
formation can be externally conveyed. As such, it has been up to
their employees (e.g., through internal processes of budgeting and
forecasting). However, in recent years, advances in technology have
resulted in social media platforms through which employee opin-
ions can be directly shared and aggregated. In this paper, we assess
how the opinions that rank-and-file employees voluntarily and
anonymously submit on social media relate to information that will
later be conveyed in their employers’ voluntary disclosures and
mandated accounting reports.
Glassdoor.com is a large recruiting website that, in addition to
hosting information about job postings, also hosts a social media
platform through which current and former employees can weigh
in on a variety of firm characteristics, including internal CEO
approval ratings, salary data, interview difficulty and questions,
compensation and benefits assessments, and even office photos. By

* Corresponding author.
E-mail addresses: jeffrey.hales@scheller.gatech.edu (J. Hales), robbie.moon@scheller.gatech.edu (J.R. Moon), swensola@uwm.edu (L.A. Swenson).
1
See Verrecchia (2001) for a broad review of research on disclosure.

https://doi.org/10.1016/j.aos.2018.04.004
0361-3682/© 2018 Elsevier Ltd. All rights reserved.
 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108
2017, Glassdoor reported hosting over 8 million individual
employer reviews submitted by a current or former employee.2
From this data, we construct a sample of 158,352 reviews repre-
senting 1265 of the S&P 1500 firms.
The structure of Glassdoor reviews allows current and former
employees to provide quantitative (discrete) assessments of 9
different attributes. Most relevant to us, employees explicitly assess
their opinion of the firm’s outlook on a three-point scale (positive,
neutral, negative). Employees rate six other attributes on a five-
point scale: work-life balance, culture and values, career opportu-
nities, compensation and benefits, senior management, and an
overall rating, and CEO approval is assessed on a three-point scale
(approve, neutral, disapprove). Finally, each review also includes a
“yes/no” question about whether reviewers would recommend
their employers to a friend.3
As mentioned, our primary focus is on the employee’s opinion of
firm outlook. However, since many of the employees’ responses
likely reflect the same latent constructs, we also use factor analysis,
which collapses the nine attributes into a smaller set of variables, to
identify which factor most likely reflects employees’ opinions of the
firm’s future prospects. Our factor analysis suggests three basic
dimensions. The most significant factor captures the employee’s
general feelings about the firm, which relates positively to all nine
review attributes. While less significant than the first, the second
factor indeed reflects the employee’s latent beliefs about the firm’s
future, which relates most strongly to explicit outlook, recom-
mendation to friend, and CEO approval. Finally, our analysis iden-
tifies a third significant factor that we interpret as opinions about
job satisfaction and retention likelihood, which relates most
strongly to the employee’s assessment of career opportunities,
opinion of compensation and benefits, and overall rating of the
firm. In all our analyses, we utilize both the explicit measure of
business outlook that employees provided and the latent measure
of outlook we extracted from the factor analysis since each offers its
own unique advantages.
Before exploring our primary research question, we first
conduct a determinants analysis to assess the extent to which
observable firm characteristics relate to employees’ opinions and to
develop a set of control variables to use in subsequent tests. We
explore four sets of predictors: (1) basic firm characteristics, like
firm size and age, (2) recent profitability and market performance,
(3) employee growth and benefits, and (4) market risk and uncer-
tainty. The results of this analysis indicate that firm size, though not
age, relates positively to outlook. As expected, return volatility and
the book-to-market ratio both relate negatively to outlook,
consistent with uncertainty and limited growth opportunities
yielding lower forecasts of future performance. Consistent with
employees considering past performance when making an
assessment of future prospects, we find that stock returns and re-
turn on assets both relate positively to outlook. Finally, measures of
growth and spending (workforce growth and SG&A) also relate
positively to outlook.
Our primary motivation is to explore the role that rank-and-file
employees play in providing an alternative disclosure channel by
2
See https://www.glassdoor.com/about/overview.htm (accessed March 30,
2017).
3
Employees provide qualitative information as well. Specifically, employees have
the option of separately writing about the pros and cons of working for the
employer, advice to management, and an overall summary assessment. We largely
exclude this information from our analyses for a number of reasons, including the
considerable variation in length and language, which complicates objective clas-
sification of the content e especially as it relates to outlook. We leave a more
detailed and rigorous analysis of the qualitative content of the reviews to future
research.
89
utilizing social media platforms such as Glassdoor.com. In order to
do so in a meaningful way, rank-and-file employees must both hold
and reveal information about their firm’s future prospects. The first
part of this premise seems likely given employees’ completion of
job-related tasks and frequent interactions with others internal to
the organization. In addition, research suggests that investment
decisions made by employees imply that they have knowledge
relevant to future stock performance (Babenko & Sen, 2015;
Huddart & Lang, 2003). As such, publicly shared employee opinions
about the future prospects of their employer could serve as an
alternative disclosure channel about the firm, which is the focus of
our empirical tests.
For our main analyses, we examine measures derived from news
communicated by managers through future mandatory and
voluntary disclosures. We begin by assessing whether employee
outlook relates to core operating performance reflected in various
income statement line items. We choose the income statement
because most employees likely understand how their information
relates to revenue generation or cost savings within a firm. We use
multiple income statement line items because we expect employee
opinions to be more informative about core performance bench-
marks, like sales and gross margins, though we consider measures
of income as well. Our results suggest that both explicit and latent
employee outlook positively predict future sales, gross margins,
operating income, and net income, controlling for current and past
performance. We also assess whether outlook assessments predict
other more transitory items, like goodwill impairments, other
write-downs, and restructuring charges. Unlike other measures of
financial health, transitory events like goodwill impairments are
difficult for market participants to predict (Hayn & Hughes, 2006)
and are not typically priced prior to the loss announcement (Li,
Shroff, Venkataraman, & Zhang, 2011). However, our evidence
suggests that explicit and latent employee outlook relates to these
transitory events; periods where employees express more negative
outlook coincide with larger goodwill impairments, other write-
downs, and restructuring charges. Overall these two sets of ana-
lyses suggest employee opinions conveyed through social media
appear to foreshadow information about firm performance that
will be subsequently conveyed through mandatory disclosure
channels.
We next examine whether information about future perfor-
mance conveyed by employee outlook is incremental to other in-
formation available at the time of the review. Using analyst
consensus as a proxy for information available to investors at the
time of the reviews, we first examine whether measures of outlook
predict the news in future earnings surprises. Specifically, we
identify the quarterly earnings ending four to six months following
the review and compute a measure of earnings surprise using the
consensus estimates of earnings at the time of the review. Consis-
tent with our prior results, both explicit and latent measures of
outlook relate positively to earnings surprises, suggesting that
employees’ assessments of outlook are incremental to other earn-
ings news (i.e., analysts’ expectations) at the time of the review.
Second, as a sharper test of whether employee opinions can
serve, not only as an alternative disclosure channel, but also as one
that can precede (or front run) the more traditional voluntary
disclosure channels controlled by firm management, we test
whether employee opinions predict firms’ voluntary disclosures
following the review. To conduct this test, we identify quarterly and
annual management forecasts occurring in the 90 days following
the review and compute forecast news using the consensus esti-
mate of earnings for the same forecast period available at the time
of the review. We again find evidence that outlook reliably predicts
the news in management forecasts. In sum, our evidence suggests
that employees hold private information useful in predicting future
90 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108
firm performance and disclosure news.
While we posit that current employees’ unique access to private
information explains our results and attempt to control for the
broader construct of employee sentiment in our models, we
recognize that outlook may capture other dimensions of the
workforce that relate to future performance. For example, as we
have suggested, employees have an insider’s view of the business,
such that they can convey information about existing customer
relations, pending sales orders, etc., by posting a positive review of
their company. But that positive review might also reflect more
stable characteristics, like employee optimism or firm culture, that
could still translate into good or bad future firm performance
without being closely tied to timely information about customer
relations, future sales demand, or more general information that
does not require employees’ inside access to the company.
In light of these concerns, we perform two additional tests to
provide more insight into whether employee opinions about their
firm’s outlook reflect private information about their firm. First, we
examine whether former employees’ outlook assessments similarly
translate into accurate predictions of future disclosure news. We
specifically exclude former employees from our main analyses
since it is unclear on Glassdoor.com when former employees left
the firm, and the longer they have been gone, the less informative
we would expect their private information to be. When we analyze
the opinions of these former employees, we find similar associa-
tions between outlook and future mandatory and voluntary firm
disclosures as in our main tests, though the economic and statistical
significance of these results are (as expected) weaker. This set of
findings suggests that the informativeness of employee opinions
about the firm’s future outlook is more likely rooted in firm-specific
knowledge that decays with time rather than more general
knowledge. Second, we examine whether the informativeness of
current employee opinions about future performance increases as
they become more informed about the firms’ activities, which we
proxy for using employee tenure. Current employees self-report
tenure from a list of options (e.g., “less than 1 year,” “more than a
year,” “more than 3 years,” etc.). We partition the sample and
classify those with less (more) than 3 years of experience as short-
tenure (long-tenure). In most of our analyses, the relation between
outlook and our measures of future disclosure news appears
stronger in the long-tenure partition, and this difference is statis-
tically significant in several cases, again supporting our contention
that employees’ possession of private information explains our
results.
In our final additional analysis, we consider whether explicit
and latent outlook relate to future stock returns. While we believe
metrics used in our main analyses best capture firm performance,
changes in these metrics may also move subsequent stock prices,
suggesting a second order positive relation between explicit and
latent outlook and future stock returns. We find evidence that
explicit outlook predicts returns over a 6-month horizon but fail to
find statistically significant evidence that latent outlook predicts
returns over a 6-month horizon.
Our study makes several contributions to the literature. First,
the use of various websites and social media platforms as possible
sources of company-specific information is of growing interest in
accounting and finance (Miller & Skinner, 2015). Examples of
research on this broad theme includes the role of Internet bulletin
boards (Antweiler & Frank, 2004), Twitter (Bartov, Faurel, &
Mohanram, 2018; Blankespoor, Miller, & White, 2014; Bollen,
Mao, & Zeng, 2011; Jung, Naughton, Tahoun, & Wang, 2017; Lee,
Hutton, & Shu, 2015; Tang, 2017), Seeking Alpha (Campbell,
DeAngelis, & Moon, 2018; Chen, De, Hu, & Hwang, 2014), and
Estimize (Da & Huang, 2017; Jame, Johnston, Markov, & Wolfe,
2016). Particularly, Bartov et al. (2018) and Tang (2017) investi-
gate whether information disclosed through Twitter predicts a
company’s future performance, and Chen et al. (2014) and
Campbell et al. (2018) link financial commentary on Seeking Alpha
to current and future stock performance. This line of research
provides evidence that “non-traditional” news outlets reflect in-
formation relevant in assessing firms’ future prospects.
We also view our results as speaking to a much larger and more
developed literature on voluntary disclosure. While this literature
has largely modeled and analyzed voluntary disclosure as a firm-
level decision, we add to this literature by exploring the potential
for employee-dedicated, social media platforms, such as Glassdoor.
com, to serve as an outlet through which rank-and-file employees
can voluntarily communicate information about public and private
firms, presumably without managers’ consent. Miller and Skinner
(2015, p. 226e227) observe that the advent of social media has
led firms to lose “a certain amount of control of their information
environments in ways that are difficult to predict and manage.”
Consistent with this view, our evidence suggests that rank-and-file
employees are now chipping away at upper-level management’s
exclusive control over the firm’s disclosure channels. Thus, while
the choice to disclose is voluntary on the part of the employee
posting the review, from the perspective of management, these
reviews can be characterized as a type of in-voluntary (but not
mandatory) disclosure on the part of the firm.
Consequently, our results also contribute to research examining
the general question of whether employees’ own interpretation of
private information is reflected in future performance. Using option
exercises and aggregate purchases of company stock by non-
executives, prior research finds mixed results as to whether non-
executives’ decisions signal private information reflected in future
returns (Babenko & Sen, 2015; Core & Guay, 2001; Huddart & Lang,
2003). An inherent limitation of these types of studies is that they
rely on employee actions, such as stock option exercises, to proxy
for employee opinions about the firm, even though the actions
could be driven by factors unrelated to the firm (such as employee
liquidity needs). We believe that social media platforms, such as
Glassdoor.com, can serve as a good setting for examining whether
rank-and-file employees hold private information about their firms
because we can directly observe the opinions they have chosen to
share rather than having to infer an opinion about the company
from observing an action, like a stock purchase or option exercise.
Finally, we expect our results to be of interest to market in-
termediaries, like analysts or auditors. Our results suggest that
employees’ opinions expressed on Glassdoor.com provide insight-
ful information about the future, which could be useful for analysts’
recommendations. In addition, these informative reviews should be
independent of management intervention suggesting that
employee opinions expressed on Glassdoor.com could be used as
corroborating third-party audit evidence.
The remainder of the paper is organized as follows. In Section 2,
we discuss background literature and motivation related to our
research question. Section 3 describes our data, sample selection,
descriptive statistics, and results from our determinants analysis.
We describe our research design and provide results in Section 4.
Section 5 reports results from additional analyses and robustness
tests, and Section 6 concludes.
2. Prior literature & research question
An extensive line of research examines how broad workforce
characteristics, like employee sentiment, job satisfaction, or
 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108
opinions of corporate culture, relate to outcomes measuring future
performance (e.g. Edmans, 2011, 2012; Edmans, Li, & Zhang, 2014;
Huang, Li, Meschke, & Guthrie, 2015; and; Green, Huang, Wen, &
Zhou, 2017), audit risk (Huang, Masli, Meschke, & Guthrie, 2017),
or reporting quality (Ji, Rozenbaum, & Welch, 2017).4 In this study,
we add to prior work by examining whether employee opinions, as
expressed on social media platforms, are predictive of future
mandatory and voluntary disclosures at the firms where they work.
If employees both have informative opinions about the firms they
work at and are willing to share those opinions, then social media
platforms can serve as an alternative disclosure channel about
firms.
Like those higher up in an organization, rank-and-file em-
ployees almost certainly have access to private information about
the future of the firm, albeit at a much more granular level. While
executives likely receive regular reports about firm-wide in-
vestments, contracts, and other relevant information, employees’
gather their private information largely through completion of job-
related tasks. For example, an accounts receivable clerk maintains
constant communication with customers and, therefore, has inti-
mate knowledge of past due invoices and so will likely have a sense
of whether collections are improving or deteriorating, even before
those changes are reflected in reported earnings.5 Similarly, a sales
representative may complete training on a new product still in
development that the company hopes to bring to the market in the
near future. Employees may also obtain private information
through interaction with others internal to the organization,
including both other rank-and-file employees and executives.
Rank-and-file employees’ “water-cooler” talk offers a venue for
them to share private information about the firm with each other.
Much of this information, whether current or forward-looking, will
take time to make its way into and through the accounting and
reporting cycle. In addition, companies often have department,
division, or even company-wide meetings where individuals from
the executive team review the past periods’ results and propose
goals and plans for the upcoming quarter or year. Both the tone and
the content of the message delivered by the executive team provide
rank-and-file employees with private information about their
firm’s future prospects.
To our knowledge, Glassdoor.com represents the first venue in
which employees are encouraged to voluntarily convey their own
opinions about their employers, including an explicit assessment of
firm outlook. However, prior research has examined whether it is
possible to indirectly infer employee opinions by examining
observable actions and has found mixed results. On the one hand,
Core and Guay (2001) fail to find evidence that option exercises by
non-executives reflect private information about their firms. More
recently, however, Babenko and Sen (2015) use data on employee
stock purchase plans and find that firms in the top quartile of
employee stock purchases earn 10 percent higher abnormal returns
in the next 12 months than do firms in the bottom quartile. Simi-
larly, using a sample of individual option exercises covering over 50
thousand employees across seven organizations, Huddart and Lang
(2003) find that, when employees exercise fewer options, their
employer’s six-month stock return is 10% higher than when em-
ployees exercise many options.
4
Like us, Huang et al. (2015), Huang et al. (2017), Green et al. (2017), and Ji et al.
(2017) use ratings from Glassdoor.com. However, none of these studies investigate
whether employee outlook predicts future performance.
5
Whether this perspective gives any one employee a distinct information
advantage over anyone else in the firm is doubtful, given the relative size of firms
compared to the purview of a typical rank-and-file employee. However, in aggre-
gate, this information could accurately predict future corporate disclosures and
events.
91
To summarize, employees likely possess an information
advantage over outsiders due simply to their status as insiders of
the firm. Further, while research on employees’ investment de-
cisions implies that employees may convert their private informa-
tion into decision-useful information, it is unclear exactly what
triggers their investment decisions (i.e. why employees choose to
buy stock or hold options). However, with the advent of social
media, employees may now be able to more directly share their
private information by posting opinions about their firm’s future
prospects.
Despite our aforementioned reasons that employee outlook may
translate to accurate predictions of future disclosure news, we
highlight several reasons why this may not be the case. Namely,
while managers have market incentives to voluntarily disclose
expectations about future earnings (e.g., Baginski & Rakow, 2012;
Balakrishnan, Billings, Kelly, & Ljungqvist, 2014; Billings, Jennings,
& Lev, 2015), it is less clear that these benefits extend to rank-
and-file employees, given that completion of a review requires
time and effort and yields no obvious direct benefit to the rank-
and-file employee, as their opinion is neither compensated nor
likely to move market prices on its own. Further, even if they hold
substantial financial interests in the company and wish to accel-
erate price formation (Campbell et al., 2018; Pasquariello & Wang,
2018), employees could free ride off of the information dissemi-
nation of their fellow employees. Moreover, many rank-and-file
employees may hold no equity stake in the firm at all, completely
removing stock-based incentives to leave reviews. Finally, even
with the best intentions, in aggregate employees may not possess
sufficient private information or sophistication to make their
assessment of outlook informative.
3. Glassdoor data
3.1. Data and descriptive statistics
Founded in 2007, Glassdoor is a social media company that
collects and posts information about companies that is obtained
directly from current and former employees, including information
about salaries, interview experiences, and (the focus of our study)
company reviews. Glassdoor claims to be “the fastest growing jobs
and recruiting site” and now hosts over 8 million anonymous
company reviews. In this study, we examine whether the employee
opinions shared on this social media platform contain information
that will later be revealed in mandatory and voluntary corporate
disclosures.
Glassdoor’s community guidelines state that “Glassdoor strives
to be the most trusted and transparent place for today's candidate
to search for jobs and research companies” (http://help.glassdoor.
com/article/Community-Guidelines/en_US), and these guidelines
also mandate that reviews be submitted voluntarily and without
coercion by management (management may encourage employees
to submit reviews, but should not offer incentives to leave positive
feedback). Reviews are also completely anonymous, and Glassdoor
encourages employees to write clear, balanced opinions of their
employers. Additionally, Glassdoor reports using a rigorous, two-
step moderation process to detect abuse or gaming, minimizing
the likelihood that companies can unduly influence the informa-
tion disclosed by employees. Thus, Glassdoor strives to provide
useful information about firms and to prevent abuse or gaming by
reviewers or employers.
When submitting a review, individuals are asked to provide
several ratings, each measured on a five-point scale. In particular,
Glassdoor collects the respondent’s assessment of company culture
and values, compensation and benefits, work/life balance, its senior
management, and an overall rating. In addition to those ratings,
92 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108

respondents are also asked to indicate whether they would Table 1

recommend the company to a friend, to provide their opinion of the Sample attrition.

CEO, and, most relevant to our study, to give their assessment of the Total Reviews (May 2012-June 2015) 386,902
firm’s outlook. We provide the full questionnaire in Appendix A.6 less: Reviews by former employees (159,507)
One concern with the Glassdoor data (or any voluntarily less: Incomplete reviews (missing major rating components) (69,043)
Maximum Sample 158,352
completed survey) is that the person providing the review and the Unique Companies 1,265
timing of the review are non-random. Depending on the research
Table 1 presents sample attrition.
question, this could create selection concerns that complicate in-
ferences. We speculate (though we cannot verify) that many
Glassdoor reviews are prompted by companies’ encouraging em-
ployees to submit a review7 and/or employment milestones. For
1500.9 For each record identified, we then obtain Glassdoor’s firm
example, a promotion or award might encourage an employee to be
identifier, which we use to access employee reviews. We then
a steward of the firm and try to attract other high-quality in-
manually verify each link to avoid erroneously matching companies
dividuals. Alternatively, a negative experience, like being passed
in Compustat to Glassdoor. This procedure allows us to successfully
over for promotion or receiving a poor performance review, might
identify Glassdoor company profiles for 1,415 of the firms in the
cause an employee to begin looking for a new job on Glassdoor.
S&P 1500. Second, after matching Compustat identifiers to Glass-
While the nature of these events likely correlates with the tone of
door, we use two additional scripts to download and parse the
the reviews, we feel selection concerns are somewhat mitigated in
individual reviews. The first script opens the website containing the
our setting for several reasons. First, we are primarily focused on
reviews for a given company, extracts required information from
whether social media platforms can serve as an alternative
the source code for the web page, and then cycles through each
disclosure channel about firms. As such, we focus only on firms and
subsequent page in a similar fashion until all reviews for that
periods with reviews, and we do not attempt to draw inferences
company are collected. This process is repeated for the 1,415 firms
about what employees in firms with no reviews might know, if they
in our original sample. A second script then parses individual pages
were to post something about their firm. Second, our models
of reviews into normalized data (described below). For each review,
attempt to control for likely confounds. Specifically, we explicitly
we also obtain the review’s unique identifier, the date of the review,
control for observable firm, industry, and market events in our
the employee’s job title (if provided), status (current or former),
models (e.g., past performance), as we discuss shortly, and we
and tenure (if provided).
include time-period fixed effects in all models, which allows
From each review, we extract the following information (each
outlook (and other independent variables) to only explain within-
on a five-point scale): the employee’s overall rating (Overall) and
time period variation. Finally, if the selection forces are internal to
subcomponent ratings capturing the employee’s opinion of their
the firm, then they may reflect the variation we intend to captur-
employers’ culture and values (CultureValues), work/life balance
edemployees’ inside view of their firms as revealed through
(WorkLife), senior management (SeniorManagement), career op-
Glassdoor reviews.8
portunities (CareerOpps), and compensation and benefits (Comp-
We collect review data using a two-step procedure. First,
Benefits). We also extract the respondent’s assessment of the CEO
Glassdoor hosts an application program interface (API) that allows
(CEOApproval), coded as 1 (positive), 0 (neutral), or
1 (negative),
us to programmatically access certain information within their
whether he/she would recommend the company to a friend (Rec-
databases. The API provides only summary information about re-
ToFriend, coded 1 for yes or 0 for no), and his/her opinion of the
views (e.g., average overall rating, percent approving of CEO, etc.),
company’s six-month business outlook (Outlook), coded as 1
limiting its use for constructing a time series of review data.
(positive), 0 (neutral), or
1 (negative).10 We are most interested in
However, we use the API to obtain company-specific metadata that
Outlook, as it is the variable most directly related to employees’
we then use to match review data with data from CRSP and Com-
opinions that might contain information that has not yet made its
pustat. Specifically, we first identify the S&P 1500 as of 12/31/2013
way into mandatory or voluntary corporate disclosures. However,
in Compustat. We then utilize a Python Script to query Glassdoor’s
while employees respond to nine discrete dimensions, we recog-
API by company name (Compustat “conm”) for each firm in the S&P
nize that latent factors likely explain some covariation in their re-
sponses to the various questions. Therefore, we also utilize a factor
analysis to identify latent factors that may serve as an alternative
6
As noted in footnote 3, employees also provide qualitative information which measure of outlook (discussed in the next subsection).
we do not consider in this study. As discussed previously, our sample begins with employee re-
7
Glassdoor’s Community Guidelines supports employers encouraging employees views for the S&P 1500 available on Glassdoor. In total, we suc-
to share honest reviews on Glassdoor but discourages firms from offering in- cessfully download 386,902 reviews for these firms between May
centives in exchange for reviews. In fact, Glassdoor will remove reviews with evi-
2012 and June 2015. We start with May 2012 since Outlook was not
dence that employees were compensated and/or coerced into leaving a review
(http://help.glassdoor.com/article/Community-Guidelines/en_US). collected prior to that date. For our main analyses, we remove
8
For future researchers interested in questions where selection issues might be 159,507 reviews completed by former employees and 69,043
more problematic, we note that our inspection of the data suggests that employees incomplete reviews (i.e., missing at least one review component).11
providing reviews appear to provide, on average, at least somewhat balanced re-
views. For example, the mean number of words in the “pros” section is 23 words for
reviews with a positive outlook versus 19 for reviews with a negative outlook
(untabulated), implying favorable reviewers still consider negative aspects of their
10
employer. In the “cons” section, word counts are significantly higher for negative Specifically, employees are asked the following question: “In the next six
outlook reviews than positive reviews. However, the mean number of words in the months do you think your company will perform better, worse, or remain the
“cons” section for positive reviews (23) is nearly identical to the word count in the same?”
“pros” section for these reviewers, again suggesting employees try to offer a 11
Like the opinions of current employees, the opinions of former employees may
balanced view of their employers. also reveal unique information because of their inside knowledge of the firm.
9
We normalize company names before searching Glassdoor for a match by However, we are unable to determine departure dates for the former employees
replacing uncommon acronyms like “HLDGS” with “HOLDINGS” and “SOLNS” with making it difficult to compare them to current employees (or to each other).
“SOLUTIONS.” Common acronyms, like “COMP” and “INC” do not inhibit our Although we remove them from our sample for our main analyses, we analyze
searches. these reviews in a supplemental analysis described in Section 5.
 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108
Glassdoor Reviews by Month
8000
7000
6000
5000
4000
3000
2000
1000
Fig. 1. Glassdoor reviews by month.
Figure 1 presents the number of reviews by month present in our final sample, described in Table 1.
This leaves a total of 158,352 reviews representing 1,265 com-
panies. These reviews represent our starting sample for each of our
main analyses. Table 1 describes this sample attrition.
Figure 1 graphs the number of reviews in our sample for each
month between May 2012 and June 2015. Consistent with Glass-
door’s self-described rapid growth, the number of reviews per
month increases substantially over our sample period. With the
exception of the first month in our sample, current employees
completed more than 2,000 total reviews per month, and, in recent
months, that number grows to approximately 7,000.
Table 2 presents descriptive statistics for the Glassdoor ratings
data. Panel A describes the number of reviews per firm. The dis-
tribution of reviews is somewhat skewed, with a mean (median)
number of reviews per firm of 549 (204). Panel B of Table 2 displays
the sample distribution by industry designation using Fama-French
48-industry classifications. Except for Business Services (19
percent) and Retail (24 percent), most industries do not have a
disproportionate number of reviews. Further, the fact that these
industries have a higher number of reviews is consistent with
companies in these industries relying heavily on large workforces.
All of our empirical models include industry fixed effects, limiting
the variance explained by our variables of interest to within-
industry variation. However, in Section 5, we provide additional
analyses to test whether these two industries unduly influence our
main results.
Panel C of Table 2 presents descriptive statistics on review
components. The mean value of Outlook of 0.272 suggests that the
average employee posting a review has a positive opinion of the
firm’s near-term business prospects. In fact, almost 50 percent of
respondents report a positive outlook, whereas fewer than 25
percent report a negative outlook. Similarly, the mean (median)
value for Overall of 3.366 (4) on a 5.0 scale suggests that the average
employee posting a review has a generally positive opinion of the
firm. Indeed, four of the five subcomponents, WorkLife, Cultur-
eValues, CareerOpps, and CompBenefits, have mean (median) values
above (above or equal to) the midpoint of three on a five-point
scale. SeniorManagement is the only subcomponent rating with a
mean (2.952) falling below the midpoint. Despite holding an
ambivalent attitude toward senior management as a whole, em-
ployees posting a review appear to generally approve of the CEO
(mean of 0.358 for CEOApproval), and approximately 65 percent
would recommend their firm to a friend. Panel D of Table 2 presents
93
Pearson and Spearman correlations between review measures. As
expected, all components exhibit significantly positive correlations
(r between 0.36 and 0.76). These significantly positive correlations
suggest that a smaller number of latent dimensions may largely
explain the reviewer’s responses across the various parts of the
review, which we discuss more in the next section.
3.2. Factor analysis
Our primary question is whether employee opinions contain
information about future corporate disclosures and events, which
would seem to be most directly captured by our Outlook variable.
However, that one variable may be a noisy proxy for employee
opinions about the near-term prospects of their firm, and it may be
influenced by the employees’ responses to preceding questions. In
addition, employees’ overall view of their firm’s prospects is likely
to partially influence other assessments they make in the review.
Consistent with these notions, Outlook exhibits significant Pearson
correlations (r > 0.5) with Overall, RecToFriend, CEOApproval, Cul-
tureValues, CareerOpps, and SeniorManagement.12 Therefore, to
better isolate the latent constructs that drive employees’ assess-
ments of outlook, we use a factor analysis to reduce these nine
review dimensions to a smaller set of variables. A factor analysis
provides several benefits. First, it uses common variance among
review attributes to identify the latent constructs which drive the
review responses. In addition, by focusing on a smaller set of latent
constructs, we can isolate the construct that relates most strongly
to employees’ outlook assessments. From a practical perspective, it
also replaces nine, highly correlated variables with a smaller set of
less correlated measures. Note that we view this factor analysis as
neither confirmatory (i.e., we make no ex ante prediction of how
factors will load) nor purely exploratory (i.e., we expect Outlook to
be the primary latent dimension of at least one factor).
We employ principal factor scoring with promax (oblique)
rotation. We use an oblique rotation to allow extracted factors to be
correlated, which we believe to be more theoretically valid.13 For
example, opinions of senior management are likely related to both
12
The remaining two review questions, Worklife and CompBenefits exhibit positive
correlations with Outlook of 0.38 and 0.40, respectively.
13
We also conduct an analysis with unrotated factors, as reported later in Section
5.
94 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108

Table 2
Distribution of reviews.

Panel A: Descriptive Statistics on Number of Reviews (n ¼ 158,352)

Mean Std. Dev. 25% 50% 75%

Number of Reviews per firm 549.207 761.484 57 204 711

Panel B: Distribution of Reviews by Industry

Fama-French Industry (48-classification) Total Reviews % of Total Reviews Average Reviews per Firm Median Number of Reviews per Firm

Agriculture 151 0.10% 74 74

Food Products 1,369 0.86% 62 41
Candy & Soda 675 0.43% 228 211
Beer & Liquor 917 0.58% 386 382
Tobacco Products 276 0.17% 43 41
Recreation 151 0.10% 31 26
Entertainment 732 0.46% 61 45
Printing and Publishing 311 0.20% 37 30
Consumer Goods 1,908 1.20% 175 88
Apparel 1,896 1.20% 79 60
Healthcare 1,002 0.63% 57 42
Medical Equipment 1,608 1.02% 57 39
Pharmaceutical Products 2,923 1.85% 95 64
Chemicals 1,989 1.26% 168 96
Rubber and Plastic Products 47 0.03% 10 8
Textiles 102 0.06% 21 20
Construction Materials 326 0.21% 29 12
Construction 691 0.44% 69 42
Steel Works Etc 306 0.19% 19 12
Machinery 2,663 1.68% 79 53
Electrical Equipment 656 0.41% 82 63
Automobiles and Trucks 1,022 0.65% 121 82
Aircraft 1,356 0.86% 285 206
Shipbuilding, Railroad Equipment 33 0.02% 12 11
Defense 853 0.54% 349 343
Non-Metallic and Industrial Metal Mining 103 0.07% 16 10
Coal 47 0.03% 7 7
Petroleum and Natural Gas 2,395 1.51% 117 69

Utilities 1,130 0.71% 30 19

Communication 5,894 3.72% 540 275
Personal Services 832 0.53% 77 50
Business Services 29,979 18.93% 727 243
Computers 9,755 6.16% 890 406
Electronic Equipment 8,322 5.26% 348 182
Measuring and Control Equipment 1,752 1.11% 160 82
Business Supplies 870 0.55% 59 46
Shipping Containers 165 0.10% 22 18
Transportation 3,464 2.19% 248 131
Wholesale 2,169 1.37% 71 44
Retail 37,974 23.98% 829 505
Restaurants, Hotels, Motels 6,646 4.20% 531 266
Banking 12,196 7.70% 728 488
Insurance 5,205 3.29% 178 104
Real Estate 487 0.31% 90 79
Trading 3,465 2.19% 114 58
Other 1,483 0.94% 261 138
Missing SIC 56 0.04% 19 17

Total 158,352 100.00% 185 60

Panel C: Descriptive Statistics on Review Components (n ¼ 158,352)

Variables Mean Std. Dev. 25% 50% 75%

Outlook 0.272 0.764 0 0 1

Overall 3.366 1.207 3 4 4
RecToFriend 0.650 0.477 0 1 1
CEOApproval 0.358 0.757 0 1 1
WorkLife 3.270 1.281 2 3 4
CultureValues 3.378 1.338 2 4 5
CareerOpps 3.195 1.246 2 3 4
CompBenefits 3.291 1.183 3 3 4
SeniorManagement 2.952 1.296 2 3 4

Panel D: Correlations on Review Components (Spearman Above/Pearson Below) (n ¼ 158,352)

Overall Outlook RecToFriend CEOApproval WorkLife CultureValues CareerOpps CompBenefits SeniorManagement

Overall 0.600 0.725 0.576 0.572 0.748 0.717 0.595 0.756

Outlook 0.604 0.560 0.510 0.364 0.532 0.523 0.389 0.565
RecToFriend 0.732 0.568 0.546 0.475 0.635 0.593 0.462 0.637
 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108 95

Table 2 (continued )

Panel D: Correlations on Review Components (Spearman Above/Pearson Below) (n ¼ 158,352)

Overall Outlook RecToFriend CEOApproval WorkLife CultureValues CareerOpps CompBenefits SeniorManagement

CEOApproval 0.588 0.516 0.553 0.366 0.553 0.478 0.396 0.565

WorkLife 0.587 0.376 0.490 0.378 0.541 0.427 0.389 0.542
CultureValues 0.760 0.541 0.654 0.566 0.553 0.615 0.472 0.722
CareerOpps 0.723 0.525 0.599 0.485 0.440 0.625 0.531 0.644
CompBenefits 0.601 0.396 0.475 0.403 0.401 0.483 0.542 0.488
SeniorManagement 0.758 0.567 0.639 0.570 0.551 0.726 0.650 0.497

Table 2 presents descriptive statistics and distributional data for the sample of reviews used in this study. Panel A presents descriptive statistics on number of reviews per firm.
All untransformed continuous variables are winsorized at the 1st and 99th percentile of their distributions. Panel B presents distributions of the reviews by industry. Panel C
presents descriptive statistics on review components and Panel D presents correlations among review components; correlations with significance levels <0.05 are in bold. All
variable definitions appear in Appendix B.

Table 3
Factor analysis.

Variables Employees’ Overall Opinion of the Firm Employees’ Outlook of Future Firm Employees’ Expectations about Personal Growth
Performance Opportunities

(SentimentFactor) (OutlookFactor) (RetentionFactor)

Outlook 0.5190 0.4227 0.2103

Overall 0.7962 0.2664 0.3601
RecToFriend 0.6615 0.3649 0.2406
CEOApproval 0.5374 0.4068 0.1609
WorkLife 0.6377 0.0684 0.1305
CultureValues 0.7775 0.2628 0.1696
CareerOpps 0.6193 0.2500 0.4150
CompBenefits 0.5117 0.1345 0.3863
SeniorManagement 0.7647 0.2792 0.2072

Eigenvalue 5.1029 0.1229 0.0937

% Variance 0.7876 0.1588 0.1359
Explained

Table 3 presents results from a factor analysis conducted on the nine review components. Reported factor loadings are based on principle factoring with promax (oblique)
rotation. All variable definitions can be found in Appendix B.

job satisfaction and expectations about future performance. We use
parallel analysis to determine the number of factors to retain (Horn,
1965; Humphreys & Ilgen, 1969; Humphreys & Montanelli, 1975;
Montanelli & Humphreys, 1976). In parallel analysis, eigenvalues
obtained from the factor analysis are compared to eigenvalues from
simulated, random data with similar distributional properties. The
rationale with parallel analysis is that nontrivial components from
our data with valid underlying factor structure should have larger
eigenvalues than parallel components derived from random data
(Ford, MacCallum, & Tait, 1986; Hayton, Allen, & Scarpello, 2004;
Lautenschlager, 1989). This analysis is repeated a number of times
(in our case, 100), and the mean eigenvalue for each factor from the
simulated data is compared to the eigenvalue from the factor
analysis. Factors with eigenvalues greater than factors from simu-
lated data are considered significant.14 Thus, the critical value for
factor “significance” varies by factor, with more (less) significant
factors having higher (lower) eigenvalue “hurdles” from the
14
Other common factor extraction methods are Kaiser’s test and Catell’s scree
test. These tests are often selected due to their ease in implementation (Velicer,
Eaton, & Fava, 2000) but have been criticized for both overestimating and under-
estimating the number of relevant factors (Cattell & Jaspers, 1967; Cattell &
Vogelmann, 1977; Hakstian, Rogers, & Cattell, 1982; Hubbard & Allen, 1987; Linn,
1968; Tucker, Koopman, & Linn, 1969; Zwick & Velicer, 1982, 1986). Thus, more
recent research suggests that parallel analysis, which is non-parametric and im-
poses no ex ante assumptions, produces a more accurate method for identifying the
true number of factors (Lance, Butts, & Michels, 2006; Velicer et al., 2000, p. 67).
simulated data.
This procedure results in three factors, and we present factor
loadings in Table 3. When discussing factor loadings, we focus on
loadings greater than 0.3 for brevity, though all loadings are used
for estimating each of the three factors. As shown in column 1, the
first principal factor, has an eigenvalue exceeding 5 and is, by far,
the most significant factor in explaining review content. It relates
positively to all nine review components, consistent with the em-
ployee’s general opinion and feelings towards the company
explaining a large portion of his or her review. We label this vari-
able SentimentFactor. Our primary motivation for conducting this
analysis is to identify a second measure that reflects employees’
latent outlook, or their opinions about the firm’s future prospects
revealed in the complete review. The pattern of loadings in column
2 are consistent with this intent. Specifically, Outlook, RecToFriend,
and CEOApproval are the main contributors to the second factor.
Each of these factors logically relates to the company’s future
prospects. In contrast, WorkLife and CompBenefits, which reflects
attributes that are costly for a firm to maintain, exhibit much
smaller loadings (0.07 and 0.13, respectively). Thus, we label this
variable OutlookFactor and note that, as expected, it is highly
correlated with Outlook (r ¼ 0.83). Finally, as reported in column 3,
one additional factor has an eigenvalue that exceeds its randomly
generated counterpart. Overall, CareerOpps, and CompBenefits
contribute the most to this third factor. We interpret this third
factor as relating to employees’ view of personal growth
96 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108

opportunities and their satisfaction with compensation. Since these
views come from current employees, we label this variable
RetentionFactor.
Note that while SentimentFactor is clearly the most significant
factor in explaining the content of the reviews, this does not imply
that the remaining factors are insignificant nor that they capture
similar constructs to employee sentiment. Further, if OutlookFactor
reflects a meaningless (insignificant) combination of review com-
ponents, this biases against rejection of the null in our tests when
using it as a measure of latent outlook.
3.3. Determinants analysis
Before moving to our primary research question, we consider
firm and market characteristics associated with our two measures
of employees’ outlook assessments (Outlook and OutlookFactor). We
conduct this analysis primarily to assess the extent to which
observable firm characteristics relate to employees’ opinions and to
develop a set of control variables to use in future tests. We focus on
four sets of attributes: 1) general firm characteristics, 2) recent
profitability and market performance, 3) employee growth and
well-being, and 4) market risk and uncertainty. Given Outlook and
OutlookFactor are both measured at the review level (which occur
sporadically over our sample period), whereas other data, like firm
and analyst characteristics, are only available at fixed quarterly or
monthly intervals, we collapse our data to the firm-quarter level by
computing the mean of each measure by firm and quarter.15 Using
this firm-quarter dataset, we estimate the following model:
factors are a result of employing promax (oblique) rotation.
Results from estimating (1) are presented in Panel C of Table 4.
Column 1 (2) presents results using Outlook (OutlookFactor) as the
dependent variable. We find that firm size relates positively to both
measures (t-statistics between 5.3 and 5.9). That is, employees of
larger firms express a more positive outlook for their firms than do
employees of smaller firms. We also find that measures of past
performance (ROA, Returns) relate positively to both of our mea-
sures of outlook (t-statistics between 6.6 and 10.5). For employee
growth and well-being, both proxies (SGA and EmpGrowth) exhibit
significant, positive associations (t-statistics between 1.6 and 8.0)
with the outlook measures. These results are consistent with
growing workforces and higher wages leading to more favorable
employee opinions and expectations of the firm. Conditioning on
size, firm age does not exhibit a significant association with either
outlook measure. RetVol relates negatively to outlook (t-statistics
between
2.0 and
2.7), suggesting uncertainty depresses
employee outlook. BTM also exhibits negative associations with the
outlook measures (t-statistics between
6.0 and
7.0), suggesting
that employee outlook is lower for firms that face fewer growth
opportunities. Lastly, n relates negatively to outlook suggesting
periods of higher (lower) review intensity correspond to less
(more) favorable outlook.
Overall, employee outlook relates in predictable ways to
observable indicators of how the firm will perform in future pe-
riods. However, we highlight that adjusted R2 values in Panel C of
Table 4 are relatively low (between 8 and 9 percent), suggesting
that other, perhaps private, information likely explains a large
portion of employees' opinion of outlook.

Outlook or OutlookFactor ¼ a0 þ a1Size þ a2Age þ a3ROA þ a4Returns

4. Empirical results
þ a5SGA þ a6EmpGrowth þ a7RetVol þ a8BTM þ a9n þ e (1)
For our main analyses, we begin with mandatory reports and
For firm characteristics, we include the natural log of assets
examine whether employee outlook is predictive of a company’s
(Size) and firm age (Age).16 We make no predictions as to how
subsequently reported accounting performance. We supplement
employee outlook relates to these characteristics. For profitability
these tests in Section 4.2 by examining whether employees’ posted
and performance, we include return on assets (ROA) and market
opinions about firm outlook are also informative about what
returns (Returns). We expect employees to lean, on average, toward
management will subsequently say about future firm performance.
momentum expectations and express a more (less) positive outlook
Specifically, we examine whether employee outlook relates to
for the firm when their firm has been performing better (worse).
earnings news in future voluntary disclosures. Note that we
For employee growth and well-being, we include the ratio of SG&A
construct the sample used in each test slightly differently
expenses to sales (SGA), a rough proxy for wages and benefits paid
depending on the periodicity of the dependent variable (e.g., firm-
to employees, and growth in the workforce (EmpGrowth). We
quarter, firm-year). All regressions include time (e.g., quarter,
expect both of these measures to also relate positively to employee
month) and industry fixed effects and two-way standard error
outlook. We include return volatility (RetVol) to proxy for general
clustering consistent with the dependent variable’s unit of measure
uncertainty about the firm’s prospects and BTM as an inverse proxy
(e.g., firm and calendar quarter, firm and calendar month, etc.).
for perceived growth opportunities. We expect both measures to
relate negatively to employee outlook. Finally, we control for the
4.1. Relation between employee outlook and information in
variation in the number of reviews across firm-quarters by
mandatory reports
including the natural log of the number of Glassdoor reviews per
firm-quarter (n) and make no prediction as to how this variable
4.1.1. Relation between employee outlook and fundamental
relates to employee outlook.
performance
We present descriptive statistics on our variables used in our
We begin by assessing how employee outlook relates to mea-
determinants analysis and correlations between these variables in
sures of future fundamental performance appearing in the income
Panels A and B, respectively, of Table 4. Descriptive statistics sug-
statement. We expect employee outlook to be more strongly
gest the average firm in our sample is large, with over 4 billion in
related to items “further up the income statement” because rank-
assets (exp(8.326)), is well-capitalized (BTM of 0.433), and experi-
and-file employees likely have more information about changes
ences positive growth in their workforce (EmpGrowth of 5.7
in sales or COGS than about variations in effective tax rates and
percent). Correlations between factors are significantly positive and
interest agreements, but we consider a wide range of measures. For
range from 0.35 to 0.63. The significant correlations between our
these tests, we utilize a firm-quarter unit of measure to correspond
with firms’ quarterly income statements. Specifically, we test
15
whether Outlook relates to future firm performance using future
In all tests, we collapse reviews down to a firm-time level (e.g., firm-quarter,
firm-month, firm-day). We discuss our rationale for selecting each unit of mea-
sales (Sale), gross margin (GM), operating income (OpInc), and in-
sure when describing our tests. come before extraordinary items (Income). For each measure, we
16
We provide exact variable definitions in Appendix B. compute cumulative performance over the two quarters ending
 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108 97

Table 4
Determinants data and regression results.

Panel A: Descriptive Statistics for Review Data and Determinants (n ¼ 9472)

Variable Mean Std. Dev. 25% 50% 75%

Outlook 0.2511 0.5305 0.0000 0.2727 0.6364

OutlookFactor
0.0249 0.3619
0.2402 0.0111 0.2274
SentimentFactor
0.0586 0.6377
0.4403
0.0176 0.3714
RetentionFactor
0.0399 0.3281
0.2324
0.0258 0.1673
Size 8.3262 1.5946 7.1391 8.2126 9.3734
Age 9.0021 0.8153 8.6197 9.0076 9.6166
ROA 0.0147 0.0195 0.0052 0.0137 0.0236
Returns 0.0412 0.1211
0.0284 0.0395 0.1093
SGA 0.2636 0.1677 0.1263 0.2388 0.3694
EmpGrowth 0.0565 0.1396
0.0122 0.0339 0.0983
RetVol 0.2754 0.1127 0.1970 0.2490 0.3239
BTM 0.4328 0.2928 0.2257 0.3715 0.5735
n 1.8861 1.0988 1.0986 1.6094 2.4849

Panel B: Correlations (Spearman Above/Pearson Below) (n ¼ 9472)

Variable Outlook OutlookFactor SentimentFactor RetentionFactor Size Age ROA Returns SGA EmpGrowth RetVol BTM n

Outlook 0.839 0.649 0.450 0.047 ¡0.021 0.155 0.077 0.011 0.197 ¡0.064 ¡0.163 ¡0.041
OutlookFactor 0.839 0.627 0.377 0.055
0.013 0.166 0.061
0.010 0.196 ¡0.058 ¡0.157 ¡0.028
SentimentFactor 0.652 0.596 0.542 0.096 0.003 0.097
0.004 0.049 0.073 ¡0.040 ¡0.090 0.028
RetentionFactor 0.451 0.352 0.566 0.167 0.026 0.063
0.012 ¡0.023 0.068 ¡0.055 ¡0.094 0.061
Size 0.063 0.075 0.110 0.162 0.384 ¡0.050 0.027 ¡0.212 ¡0.118 ¡0.458 0.030 0.410
Age
0.011 0.005 0.012 0.019 0.337 0.008 0.002 ¡0.123 ¡0.250 ¡0.315 0.053 0.134
ROA 0.155 0.172 0.082 0.049 ¡0.029
0.017 0.056 ¡0.061 0.159 ¡0.185 ¡0.464 0.109
Returns 0.084 0.067
0.008
0.016 0.009
0.010 0.061
0.009
0.008 ¡0.085 ¡0.080 0.032
SGA 0.026
0.005 0.065
0.001 ¡0.207 ¡0.123 ¡0.117
0.010 0.043 0.122 ¡0.123 0.065
EmpGrowth 0.134 0.133 0.029 0.035 ¡0.090 ¡0.232 0.089
0.017 0.050 0.111 ¡0.178 ¡0.040
RetVol ¡0.099 ¡0.096 ¡0.050 ¡0.054 ¡0.394 ¡0.257 ¡0.205 ¡0.072 0.140 0.095 0.173 ¡0.198
BTM ¡0.173 ¡0.161 ¡0.084 ¡0.087 0.068 0.050 ¡0.374 ¡0.093 ¡0.100 ¡0.132 0.193 ¡0.176
n 0.014 0.031 0.065 0.078 0.482 0.148 0.083 0.016 0.051 ¡0.055 ¡0.184 ¡0.142

Panel C: Determinants Regression

VARIABLES (1) (2)

Outlook OutlookFactor

Size 0.039*** 0.027***

(5.87) (5.35)
Age
0.009 0.000
(-0.79) (0.03)
ROA 2.873*** 2.232***
(10.09) (10.47)
Returns 0.340*** 0.197***
(7.20) (6.66)
SGA 0.218*** 0.072*
(3.44) (1.69)
EmpGrowth 0.394*** 0.279***
(7.95) (7.45)
RetVol
0.222***
0.126**
(
2.65) (
2.09)
BTM
0.223***
0.142***
(-6.97) (-6.02)
n
0.036***
0.020***
(
4.78) (
3.50)

Observations 9,472 9,472

Adjusted R2 0.091 0.085

Table 4 presents descriptive statistics, correlations, and results from a determinants analysis. All variable definitions appear in Appendix B, and all untransformed continuous
variables are winsorized at the 1st and 99th percentile of their distributions. Panel A presents descriptive statistics for review factors and determinants. Panel B presents
correlations among variables in Panel A; correlations with significance levels <0.05 are in bold. Panel C presents results from regressing outlook variables on determinants.
Industry fixed effects, based on the Fama-French 48 classifications, and quarter-end fixed effects are included. Standards errors are robust to heteroskedasticity and clustering
at the firm-and quarter-end level. ***, **, and * indicate significance at the 0.01, 0.05, and 0.10 level.

4e6 months after the review to coincide with the employee’s as- performance metrics by assets in period t. We regress each per-
sessments of performance over the next 6 months.17 We scale all formance metric on employee outlook (Outlook or OutlookFactor),
our set of determinants identified in Section 3.3, and firm perfor-
mance over the four quarters preceding those used in the
17
We elect to use future levels of performance and control for current and past
performance. We find similar results if we instead use seasonally adjusted changes
in performance as our dependent variable and control for current and recent past
performance.
98 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108

Table 5
Descriptive statistics for variables of interest and controls.

Variable Measurement Base n Mean Std. Dev. 25% 50% 75%

Sale Firm-quarter 9,490 0.5315 0.3969 0.2749 0.4219 0.6748

GM Firm-quarter 9,427 0.1945 0.1205 0.1131 0.1729 0.2514
OpInc Firm-quarter 9,375 0.0537 0.0435 0.0273 0.0479 0.0729
Income Firm-quarter 9,499 0.0293 0.0384 0.0104 0.0274 0.0474
GWImp Firm-quarter 8,270 0.1391 0.7546 0.0000 0.0000 0.0000
Writedown Firm-quarter 9,499 0.0449 0.2145 0.0000 0.0000 0.0000
Restructure Firm-quarter 9,499 0.1849 0.3771 0.0000 0.0000 0.2007
Following Firm-month 19,195 2.4832 0.6040 2.0794 2.5649 2.9444
FcstDisp Firm-month 19,195 0.0509 0.0530 0.0200 0.0300 0.0600
EarningsSurprise Firm-month 19,195
0.1044 0.5705
0.1854
0.0151 0.0902
MgmtFcst_Horizon Forecast (Firm-day) 5,821 4.9560 4.6616 1.9108 3.5478 6.2775
QTRInd Forecast (Firm-day) 5,821 0.2271 0.4190 0.0000 0.0000 0.0000
MgmtFcstNews Forecast (Firm-day) 5,821
0.0010 0.0050
0.0016
0.0003 0.0004

Table 5 presents descriptive statistics for variables of interest and other control variables not previously presented. All variable definitions appear in Appendix B. All un-
transformed continuous variables are winsorized at the 1st and 99th percentile of their distributions.

dependent variable.18 Because future performance is measured
using quarterly data, we collapse all independent variables to the
firm-quarter level. Descriptive statistics for our dependent vari-
ables are presented in Table 5.
Regression results from our future performance tests are re-
ported in Table 6. Besides variables listed, we include firm-quarter
and Fama French 48 industry fixed effects and cluster standard
errors by firm and quarter. For each dependent variable (both here
and in future analyses), we estimate three specifications: one using
Outlook only, one using Outlook and controlling for other review
responses (e.g., Overall, RecToFriend, etc.), and one using the factors
derived from our factor analysis. Columns 1 through 3 (4 through 6,
7 through 9, 10 through 12) present results using Sale (GM, OpInc,
Income) in the two quarters following the review to measure future
performance. Focusing on the first column for each performance
measure (columns 1, 4, 7, and 10), we find positive and highly
significant coefficients (t-statistics between 5.6 and 8.3) on Outlook
in all specifications. In the second column for each measure (col-
umns 2, 5, 8, and 11), we introduce the other review components.
As shown, we continue to observe highly significant coefficients on
Outlook (t-statistics between 4.8 and 6.3) and the magnitudes of
these coefficients actually increase after considering other com-
ponents. The final column for each performance metric uses the
factors derived from our factor analysis. Like Outlook, OutlookFactor
exhibits similar positive and highly significant associations with
each measure of future firm performance (t-statistics between 4.1
and 6.3). These effects represent economically significant im-
provements as well. For instance, moving from negative to positive
outlook (
1 to 1) corresponds to an increase in Sale of up to 2.4
percent of assets, or 4.5 percent of the sample mean for Sale.
Together, these results suggest that the opinions employees share
on Glassdoor about the firm’s business outlook reveal information
about their firms’ future prospects as measured on the income
statement.
Turning to the other variables in the regressions, we see that
other aspects of employee reviews fail to consistently translate to
future performance. CompBenefits (Overall) exhibits a significantly
18
To illustrate, for a calendar-year firm, we regress performance over Q1 and Q2
of 2014 on employee outlook from reviews submitted in December of 2013 through
February of 2014, controlling for performance in Q1, Q2, Q3, and Q4 of 2013. This
design ensures that earnings announcements for a given quarter occur well after
the end of each review accumulation period. Note that our results are robust to
using only the second quarter of this 6-month window (Q2 2014 in the previous
example), which does not overlap at all with the review period, or to using only
current period performance (Q1 2014 in this example).
positive (negative) association with Sale in column 2 but fails to
load in any other specification. SeniorManagement also loads posi-
tively in column 11 but is insignificant in other models. In column 8,
we actually observe negative (though only marginally significant)
associations between OpInc and both WorkLife and CareerOpps. We
observe a similar mixed pattern of results with the other factors
derived in the factor analysis. Only SentimentFactor exhibits a sig-
nificant association (column 6), relating negatively to GM. Thus,
employee outlook, measured explicitly or through a latent factor
analysis, represents the only review component that consistently
translates to accurate predictions of future performance.
4.1.2. Relation between employee outlook and transitory expenses
While the prior section focuses on the relation between
employee outlook and several fundamental income statement
measures, including earnings, prior research suggests transitory,
one-time expenses, like goodwill impairments, inventory write-
downs, or restructuring charges, are difficult for investors to pre-
dict (Hayn & Hughes, 2006) and represent significant market news
events (Ayres, Campbell, Chyz, & Shipman, 2017; Bens, Heltzer, &
Segal, 2011; Li et al., 2011). To the extent that these transactions
reflect lagged information about the condition of firm assets or
expected losses related to reorganizations, rank-and-file employees
might hold information about these events while also facing fewer
direct incentives to withhold such information compared to upper-
level management. We, therefore, test whether rank-and-file em-
ployees’ assessments of outlook carry information about future
one-time income statement charges.
For this test, we define three measures of transitory expenses.
First, we consider goodwill impairments, GWImp, which equals the
total goodwill impairment in Compustat scaled by assets and
multiplied by
1 (so that larger values correspond to larger im-
pairments); note GWImp is set to missing for firms with no goodwill
on the balance sheet. Second, we consider write-downs of assets
other than goodwill (Writedown), again scaled by assets and
multiplied by
1. Finally, we examine restructuring charges
(Restructure). Unlike the prior two measures, restructuring charges
often capture expenses associated with planned organizational
changes, such as plant closures or layoffs. Like GWImp and Write-
down, Restructure is scaled by assets and multiplied by
1. As in the
prior section, we collapse the review data to the firm-quarter level
to correspond to the same measurement window as these events.
We use the same time line as our fundamental performance tests
and construct these measures over the two quarters following the
review (see footnote 18 for illustration). Since these three variables
Table 6
Future performance.

Variables (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)

Sale Sale Sale GM GM GM OpInc OpInc OpInc Income Income Income

Outlook 0.009*** 0.012*** 0.004*** 0.006*** 0.003*** 0.004*** 0.003*** 0.005***

(5.63) (4.87) (8.30) (5.50) (8.18) (6.32) (6.20) (5.83)
OutlookFactor 0.012*** 0.007*** 0.004*** 0.005***
(4.08) (6.26) (5.90) (5.52)
SentimentFactor
0.002
0.001**
0.000
0.000
(
1.09) (
2.09) (
1.13) (
0.48)
RetentionFactor 0.005 0.001
0.001
0.001
(1.15) (0.69) (
0.81) (
1.00)
Size
0.007***
0.008***
0.007***
0.003***
0.003***
0.003***
0.001***
0.001***
0.001***
0.001***
0.001***
0.001***
(
4.99) (
5.15) (
5.00) (
6.19) (
6.48) (
6.40) (
3.67) (
3.70) (
3.57) (
2.96) (
2.87) (
2.82)
Age 0.000 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.001 0.001 0.001
(0.27) (0.38) (0.25) (0.52) (0.57) (0.45) (0.97) (0.96) (0.87) (1.30) (1.23) (1.17)
ROA
0.736***
0.734***
0.732***
0.214***
0.214***
0.212*** 0.010 0.009 0.009

J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108

(-6.97) (-6.95) (-6.94) (-4.43) (-4.41) (-4.39) (0.28) (0.25) (0.27)
Returns 0.049*** 0.049*** 0.050*** 0.032*** 0.031*** 0.032*** 0.029*** 0.029*** 0.029*** 0.027*** 0.027*** 0.028***
(5.51) (5.36) (5.57) (7.87) (7.69) (7.83) (7.90) (7.84) (7.88) (7.11) (6.93) (7.07)
SGA
0.006
0.006
0.004 0.011* 0.012** 0.012**
0.001
0.001
0.001
0.005**
0.004*
0.005*
(
0.61) (
0.66) (
0.47) (1.91) (2.02) (2.11) (
0.45) (
0.32) (
0.23) (
2.03) (
1.74) (
1.72)
EmpGrowth
0.045***
0.046***
0.045***
0.015***
0.016***
0.015***
0.008***
0.008***
0.008***
0.006*
0.007*
0.006*
(
3.89) (
4.00) (
3.94) (
3.73) (
3.83) (
3.75) (
3.21) (
3.29) (
3.14) (
1.71) (
1.85) (
1.70)
RetVol
0.040**
0.041**
0.041**
0.021***
0.020***
0.021***
0.020***
0.019***
0.020***
0.022***
0.022***
0.022***
(
2.42) (
2.45) (
2.43) (
3.69) (
3.72) (
3.69) (
4.67) (
4.72) (
4.66) (
3.48) (
3.47) (
3.47)
BTM
0.024***
0.023***
0.024***
0.008***
0.008***
0.008***
0.007***
0.007***
0.008***
0.014***
0.014***
0.014***
(
3.92) (
3.81) (
4.00) (
4.36) (
4.28) (
4.45) (
5.11) (
5.07) (
5.26) (
8.21) (
8.12) (
8.27)
n 0.004*** 0.004*** 0.004*** 0.002*** 0.002*** 0.002*** 0.001** 0.001** 0.001** 0.001** 0.001*** 0.001**
(2.91) (3.13) (2.85) (4.27) (4.47) (4.30) (2.35) (2.53) (2.31) (2.58) (2.85) (2.55)
Performancet 0.605*** 0.604*** 0.604*** 0.508*** 0.507*** 0.508*** 0.314*** 0.314*** 0.315*** 0.328*** 0.326*** 0.328***
(6.87) (6.86) (6.85) (5.27) (5.26) (5.26) (4.56) (4.58) (4.57) (5.33) (5.30) (5.32)
Performancet-1 0.115** 0.116** 0.116** 0.026 0.026 0.027
0.012
0.013
0.011 0.149*** 0.150*** 0.151***
(2.42) (2.43) (2.44) (0.51) (0.51) (0.54) (
0.31) (
0.33) (
0.29) (4.65) (4.65) (4.68)
Performancet-2 0.712*** 0.711*** 0.712*** 0.801*** 0.800*** 0.800*** 0.827*** 0.828*** 0.827*** 0.595*** 0.595*** 0.595***
(11.45) (11.43) (11.44) (13.70) (13.69) (13.69) (18.60) (18.69) (18.58) (11.97) (11.88) (11.95)
Performancet-3 0.512*** 0.513*** 0.511*** 0.567*** 0.568*** 0.566*** 0.634*** 0.634*** 0.632*** 0.496*** 0.497*** 0.496***
(6.47) (6.48) (6.47) (7.29) (7.31) (7.27) (10.05) (10.07) (10.02) (10.52) (10.59) (10.51)
Overall
0.004*
0.001
0.000
0.000
(
1.82) (
0.98) (
0.96) (
0.57)
RecToFriend 0.001
0.001
0.001
0.001
(0.24) (
0.33) (
0.45) (
1.11)
CEOApproval 0.001 0.001
0.000 0.000
(0.28) (0.97) (
0.14) (0.35)
WorkLife
0.002
0.001
0.001*
0.000 e
(
1.23) (
1.64) (
1.65) (
0.76)
CultureValues 0.001
0.000 0.000
0.001
(0.72) (
0.11) (0.34) (
1.48)
CareerOpps
0.002
0.001
0.001*
0.001
(
0.82) (
0.95) (
1.95) (
1.32)
CompBenefits 0.004*** 0.001 0.000
0.000
(3.27) (1.46) (0.21) (
0.64)
SeniorManagement
0.000 0.001 0.001 0.002**
(
0.22) (1.07) (1.30) (2.48)

Observations 9,490 9,490 9,490 9,427 9,427 9,427 9,375 9,375 9,375 9,499 9,499 9,499
Adjusted R2 0.964 0.964 0.964 0.938 0.938 0.938 0.802 0.802 0.801 0.581 0.581 0.580

Table 6 presents results from regressing measures of fundamental performance on outlook and outlook determinants. Variable definitions appear in Appendix B. All untransformed continuous variables are winsorized at the 1st
and 99th percentile of their distributions. Industry fixed effects, based on the Fama-French 48 classifications, and quarter-end fixed effects are included. Standards errors are robust to heteroskedasticity and clustering at the firm-

99
and quarter-end-level. ***, **, and * indicate significance at the 0.01, 0.05, and 0.10 level.
100 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108

Table 7
Transitory items.

Variables (1) (2) (3) (4) (5) (6) (7) (8) (9)

GWImp GWImp GWImp Writedown Writedown Writedown Restructure Restructure Restructure

Outlook
0.093***
0.122***
0.015***
0.016**
0.048***
0.066***

(
3.61) (
3.70) (
2.75) (
2.17) (
6.73) (
9.54)
OutlookFactor
0.124***
0.026***
0.065***
(
3.72) (
3.20) (
7.34)
SentimentFactor 0.007
0.005 0.007
(0.34) (
1.02) (1.06)
RetentionFactor 0.016 0.025*** 0.000
(0.52) (3.01) (0.04)
Size 0.009 0.005 0.008
0.006**
0.007***
0.006***
0.002
0.003
0.002
(0.83) (0.43) (0.74) (
2.54) (-2.94) (
2.78) (
0.52) (
0.83) (
0.70)
Age 0.241*** 0.246*** 0.247*** 0.021* 0.022* 0.022* 0.002 0.002 0.005
(3.47) (3.54) (3.53) (1.69) (1.78) (1.77) (0.08) (0.09) (0.23)
ROA
1.492**
1.501**
1.519**
0.146
0.128
0.130
1.100***
1.102***
1.111***
(
2.44) (
2.48) (
2.53) (
0.73) (
0.65) (
0.66) (
3.42) (
3.41) (
3.44)
Returns
0.059
0.094
0.076 0.012 0.008 0.009 0.101*** 0.092*** 0.091***
(
0.52) (
0.83) (
0.66) (0.54) (0.35) (0.40) (3.07) (2.78) (2.77)
SGA 0.052 0.063 0.046
0.022
0.022
0.022
0.047
0.043
0.048
(0.53) (0.66) (0.48) (
0.91) (
0.92) (
0.93) (
1.34) (
1.26) (
1.38)
EmpGrowth
0.018
0.015
0.017
0.004
0.003
0.003 0.008 0.009 0.009
(
1.17) (
1.01) (
1.13) (
0.72) (
0.58) (
0.61) (1.17) (1.23) (1.24)
RetVol
0.439***
0.424***
0.444***
0.036
0.033
0.035
0.176***
0.170***
0.178***
(
4.81) (
4.70) (
4.82) (
1.57) (
1.44) (
1.48) (
6.61) (
6.41) (
6.71)
BTM 0.393** 0.378** 0.395** 0.061 0.059 0.061 0.043 0.038 0.044
(2.39) (2.25) (2.38) (1.45) (1.41) (1.44) (0.86) (0.77) (0.88)
n
0.018
0.018
0.018 0.004 0.004 0.003 0.017*** 0.017*** 0.017***
(
1.07) (
1.06) (
1.06) (1.21) (1.37) (1.12) (2.99) (2.97) (3.12)
Transitory t 0.012 0.008 0.013 0.205*** 0.204*** 0.204*** 0.386*** 0.383*** 0.386***
(0.23) (0.16) (0.24) (3.11) (3.07) (3.08) (10.90) (10.80) (11.05)
Transitory t-1 0.099 0.097 0.097 0.091*** 0.093*** 0.091*** 0.210*** 0.209*** 0.211***
(1.35) (1.31) (1.32) (2.68) (2.76) (2.68) (13.77) (13.97) (14.16)
Transitory t-2 0.178*** 0.177*** 0.179*** 0.099** 0.097** 0.098** 0.254*** 0.255*** 0.255***
(2.59) (2.59) (2.63) (2.45) (2.42) (2.46) (9.40) (9.62) (9.61)
Transitory t-3 0.129** 0.129** 0.130** 0.106*** 0.108*** 0.108*** 0.201*** 0.202*** 0.202***
(1.97) (1.99) (1.98) (3.01) (3.06) (3.04) (6.51) (6.57) (6.58)
Overall 0.022 0.005 0.003
(0.82) (1.47) (0.32)
RecToFriend
0.002 0.004 0.005
(
0.04) (0.34) (0.34)
CEOApproval 0.013
0.012
0.004
(0.49) (
1.57) (
0.48)
WorkLife 0.003
0.004
0.001
(0.39) (
1.06) (
0.14)
CultureValues 0.021
0.010*** 0.008
(0.84) (-2.59) (1.36)
CareerOpps
0.010
0.004 0.001
(-0.59) (-0.72) (0.13)
CompBenefits 0.040** 0.016*** 0.008
(2.51) (5.18) (1.29)
SeniorManagement
0.039** 0.008** 0.000
(
2.54) (2.13) (0.03)

Observations 8,270 8,270 8,270 9,499 9,499 9,499 9,499 9,499 9,499
Adjusted R2 0.057 0.058 0.056 0.040 0.043 0.041 0.257 0.257 0.256

Table 7 presents results from regressing the likelihood and magnitude of goodwill impairments, inventory write-downs, and restructuring charges on outlook and outlook
determinants. Variable definitions appear in Appendix B. All untransformed continuous variables are winsorized at the 1st and 99th percentile of their distributions. Industry
fixed effects, based on the Fama-French 48 classifications, and quarter-end fixed effects are included. Standards errors are robust to heteroskedasticity and clustering at the
firm-and quarter-end-level. ***, **, and * indicate significance at the 0.01, 0.05, and 0.10 level.

are constructed to measure disclosure news inversely (i.e., goodwill
impairments or inventory write-downs are bad news), we expect
employee outlook to relate negatively to these measures.
Results from this test are presented in Table 7. The structure of
the table is very similar to Table 6. For each measure, we report
results using Outlook both with and without other review compo-
nents and OutlookFactor. We also include the same set of control
variables and control for values of the transitory items (Transitory)
in the four quarters preceding the window used in the dependent
variable. For GWImp (columns 1 through 3), we find highly signif-
icant, negative coefficients on Outlook and OutlookFactor (t-statis-
tics between 3.6 and 3.7), consistent with employees’
understanding how private information relates to future goodwill
impairments. In columns 4 through 6, we examine other asset
write-downs (Writedown) and continue to find evidence consistent
with our predictions. Outlook and OutlookFactor both exhibit
 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108 101

significant, negative associations with Writedown (t-statistics be- Table 8

tween 2.1 and 3.2). Finally, columns 7 through 9 present results Earnings surprises.

using Restructure. As in other specifications, both measures of Variables (1) (2) (3)
employee outlook relate significantly negatively to restructuring EarningsSurprise EarningsSurprise EarningsSurprise
charges (t-statistics between 6.7 and 9.5).
Outlook 0.046*** 0.050***
Regarding other review aspects, we again find little consistent
(4.80) (3.84)
evidence that these measures relate to future transitory items. We OutlookFactor 0.047***
do observe some evidence that RetentionFactor relates positively to (3.45)
Writedowns, implying that firms investing more in employee SentimentFactor 0.009
(0.78)
retention efforts are less likely to experience asset writedowns.
RetentionFactor
0.003
Similarly, CompBenefits relates positively to GWImp and Writedown. (
0.21)
However, most evidence related to other review components is Size 0.015 0.015 0.015
mixed or insignificant. (1.40) (1.40) (1.39)
In sum, our evidence suggests that the opinions employees Age
0.015
0.015
0.015
(
1.11) (
1.10) (
1.13)
share on Glassdoor about their firm’s business outlook contain in-
ROA 2.688*** 2.690*** 2.700***
formation about their firm’s future fundamental performance, (3.33) (3.32) (3.32)
including not only general income statement information, but also Returns 0.647*** 0.646*** 0.654***
information about future transitory, one-time expenses.19 (8.24) (8.23) (8.30)
SGA
0.087
0.086
0.084
(
1.14) (
1.14) (
1.10)
4.1.3. Relation between employee outlook and earnings surprises EmpGrowth 0.049 0.047 0.055
Recall that, when assessing their company’s business outlook, (0.64) (0.62) (0.72)
employees are asked “In the next six months do you think your RetVol
0.610***
0.609***
0.614***
(
4.35) (
4.36) (
4.36)
company will perform better, worse, or remain the same?” Because
BTM
0.039
0.039
0.041
we do not know precisely what benchmark employees might be (
0.67) (
0.66) (
0.71)
using as their benchmark when assessing their company’s business n
0.008
0.007
0.008
outlook, we next test an alternative measure of expected perfor- (
0.64) (
0.55) (
0.66)
mance derived from a clear and publicly available benchmark e Following 0.004 0.004 0.004
(0.17) (0.18) (0.17)
namely, earnings surprises. Specifically, we collapse our data to the FcstDisp
0.788***
0.792***
0.788***
firm-month level and match review data to the quarter ending four (
2.94) (
2.95) (
2.93)
to 6 months following the review. We use a firm-month definition Overall
0.019**
because consensus analyst estimates are updated once per month. (
2.41)
RecToFriend
0.025
We construct EarningsSurprise as the difference between earnings
(
1.13)
announced in this 4e6 month window and the consensus analyst CEOApproval 0.006
estimate, according to the IBES Summary File, for that period in the (0.59)
month preceding the review.20 Since reviews are submitted WorkLife 0.002
throughout the year, we use quarterly earnings announcements to (0.26)
CultureValues 0.001
avoid large fluctuations in the length of time between review and (0.13)
earnings announcement and in the scale of our earnings surprise CareerOpps
0.002
measure. We target a 4e6 month time window to coincide with the (
0.40)
6 month horizon specified in the review criteria and to ensure that CompBenefits 0.008
(1.11)
performance occurring contemporaneous to the review (which is
SeniorManagement 0.016*
reported in the 1e3 months following the review) does not (1.76)
contaminate our measure of earnings surprise.
Observations 19,195 19,195 19,195
We follow a similar research design to prior sections and regress Adjusted R2 0.104 0.105 0.104
EarningsSurprise on the two measures of employee outlook
Table 8 presents results from regressing earnings surprises on outlook and outlook
(Outlook and OutlookFactor) and the determinants discussed in
determinants. Variable definitions appear in Appendix B. All untransformed
Section 3.3. For these tests, we also control for analyst forecast continuous variables are winsorized at the 1st and 99th percentile of their distri-
following (Following) and dispersion (FsctDisp) to account for dif- butions. Industry fixed effects, based on the Fama-French 48 classifications, and
ferences in coverage characteristics of our sample firms. Results of year-month fixed effects are included. Standards errors are robust to hetero-
this analysis are presented in Table 8. Columns 1 and 2 (column 3) skedasticity and clustering at the firm-and year-month-level. ***, **, and * indicate
significance at the 0.01, 0.05, and 0.10 level.
report(s) results using Outlook (OutlookFactor) to measure em-
ployees’ assessments of the firms’ future prospects. As in previous

19
tests, we again find strong evidence that both Outlook (t-statistics
In Tables 6 and 7, we focus on mandatory disclosures four to six months in the
between 3.8 and 4.8) and OutlookFactor (t ¼ 3.5) relate positively to
future. We use this horizon because the explicit assessment of employee outlook is
for the next six months. In additional analyses, we consider a longer, twelve-month future earnings news. That is, firms whose employees report a more
horizon, and our results are qualitatively similar with the exception of our write- positive (negative) outlook are more likely to report more positive
down tests after controlling for other review components. (negative) earnings surprises, conditioned on news available at the
20
To illustrate, for reviews submitted in December 2013, we construct earnings time of the review.21 These results are economically meaningful as
surprise as the difference between earnings in Q2 of 2014 and the consensus es-
well. For instance, moving from negative to positive explicit outlook
timate for these earnings as of November 2013. Note that we would also construct
EarningsSurprise using Q2 2014 earnings for reviews submitted in October and
November of 2013, but we would use consensus estimates from September and
October, respectively. Thus, while the same period’s earnings may be used multiple
21
times, EarningsSurprise will vary by month, consistent with our decision to use a As with our findings for transitory items, we see little evidence that other as-
firm-month level of analysis. pects of the employees’ reviews relate to earnings surprises.
102 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108
corresponds to an increase in EarningsSurprise of up to 10 percent of
price.
4.2. Relation between employee outlook and voluntary disclosures
In our final set of tests, we examine whether employee reviews
convey information that will later be provided by upper-level
management through more traditional voluntary disclosure chan-
nels. In particular, we focus on management forecasts following the
review. Unlike our previous analyses, management forecasts are
themselves voluntary and thus occur sporadically over our sample
period. Therefore, we design this analysis differently from our
previous tests. Because upper-level management can voluntarily
disclose information via management forecasts at any given time
for future quarterly or annual earnings, we include both quarterly
and annual guidance in our analyses, similar to Cheng, Luo, and Yue
(2013). Specifically, for each review in our sample, we search for
earnings guidance occurring in the 90 days following the review. If
multiple forecasts exist in this window, we use the one closest to
the review, and if a quarterly and annual forecast occur on the same
day, we retain the annual forecast. We then compute the man-
agement guidance news (MgmtFcstNews) as the difference between
the forecast and the consensus earnings estimate at the time of the
review (i.e., in the most recent summary report). Finally, since each
forecast can be linked to multiple reviews, we collapse the dataset
down to the forecast level (or compute the mean of each inde-
pendent variable by forecast “firm-day”).22,23
We regress MgmtFcstNews on the same set of variables as in
Table 8. We also include MgmtFcst_Horizon, the number of days
between the forecast date and earnings announcement date, since
research suggests long-range (short-range) forecasts are more
positively (negatively) biased (i.e., “walkdown”) and QTRInd, an
indicator equaling 1 for quarterly guidance and 0 for annual guid-
ance to control for the potential difference in forecast surprise
between annual and quarterly forecasts.
Results from this regression are presented in Table 9. As in
Table 8, columns 1 and 2 (column 3) use(s) Outlook (OutlookFactor)
to measure employee assessments of firm outlook. As shown, in
column 1 (2) we find a significantly (marginally significant) positive
association between Outlook and MgmtFcstNews (t-statistics of 1.3
and 2.1, respectively) consistent with employees accurately pre-
dicting future management disclosures. Similarly, column 3 reports
a significant association between OutlookFactor and news in man-
agement forecasts (t ¼ 2.4). While not as significant as the associ-
ations between employee outlook and news in mandatory
disclosures, these results provide some evidence that employee
reviews preempt information in future management forecasts. In
other words, employee outlook provides information about how
22
To illustrate, assume management issues quarterly earnings guidance on
October 15, 2013 and February 15, 2014. Our sampling procedure would link any
reviews submitted between October 16, 2013 and February 14, 2014 to the February
15, 2014 forecast (reviews prior to but within 90 days of October 16 are associated
with that forecast). However, the value for MgmtFcstNews potentially varies by
review since we use the consensus analyst estimate as of the review date. We use
analysts’ consensus estimate at the time of the review rather than at the time the
guidance is issued (“mean_at_date” in IBES Guidance), which reflects the market’s
earnings expectations after the review, to ensure that MgmtFcstNews reflects
disclosure news as measured using information at the time reviews are submitted
to Glassdoor. For our regressions, we collapse the dataset down to the forecast-level
(or firm-day level) and use the average value of our independent variables.
23
Following prior research, we restrict the sample to point and range forecasts,
and, for range forecasts, we use the midpoint of the range to compute
MgmtFcstNews.
24
As with most of our preceding findings, we see little evidence that other aspects
of the employees’ reviews relate to management forecasts.
management forecasts will relate to the current market consensus
before management issues their forecast.24
4.3. Synthesis
Overall, our evidence suggests that the reviews employees
voluntarily share on Glassdoor reveal information that will be
disseminated in subsequent mandatory and voluntary corporate
disclosures. Specifically, using both explicit and latent measures of
employee outlook, we observe significantly positive relations be-
tween those measures and several measures of future performance
and the news in voluntary disclosures and negative relations be-
tween employee outlook and future goodwill impairments, write-
downs and restructuring charges. We view these findings as evi-
dence that, while any one review is unlikely to reveal material in-
side information, the social media platform created by Glassdoor
serves as a portal through which employees can, in aggregate, share
their inside views of the company. These results not only provide
evidence of an alternative disclosure channel, but also complement
and extend prior research that has used stock-options as a lens into
whether rank-and-file employees hold private information about
their companies (Babenko & Sen, 2015; Huddart & Lang, 2003). In
the next section, we provide a series of additional tests to provide
further insight into our main results.
5. Additional analyses
5.1. Cross-sectional tests
While our prior findings are consistent with rank-and-file em-
ployees possessing inside information, one may question whether
the informativeness of their outlook assessments results from
unique access to private information, as we argue, rather than basic
industry knowledge that could be disseminated by virtually anyone
or simply signals overall employee sentiment, which prior research
links to performance. We next perform two cross-sectional ana-
lyses to shed light on whether employee opinions about their firm’s
outlook reflect private information about their firm. First, we
consider reviews from a group of individuals that no longer have
the same type of unique access they once had e namely, former
employees.25 Second, we explore whether our results are stronger
for long-tenured employees, or employees that should be better
able to convert their private information to accurate assessments of
future performance.
5.1.1. Current vs. former employees
In our main analyses, we exclude reviews by former employees.
In this supplemental analysis, we instead use the reviews of former
employees. Specifically, we project the factor loadings reported in
Table 3 onto the subsample of former employees.26 If the infor-
mativeness of current employees’ outlook comes from their inside
view of the company, rather than, say, industry-specific or other
more general knowledge, then we should see weaker relations
25
Note that employment status and tenure are self-reported. To our knowledge,
Glassdoor cannot enforce the accuracy of these responses. However, since reviews
are anonymous it is unclear what incentives an employee would have to misreport,
and any misreporting would add noise to our partitioning process.
26
We use the factor loadings from current employees to quantify information in
the former employees’ reviews so that we can compare OutlookFactor between the
two employee groups. In an untabulated analysis, we repeat our factor analysis
using only former employees and observe a different pattern of factor loadings,
consistent with these employees responding differently to review questions (and
also confirming our choice to use current employee factor loadings to estimate the
common factors).
 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108 103

Table 9
Management forecasts.

Variables (1) MgmtFcstNews (2) MgmtFcstNews (3) MgmtFcstNews

Outlook 0.037** 0.029

(2.14) (1.33)
OutlookFactor 0.068**
(2.39)
SentimentFactor 0.007
(0.38)
RetentionFactor
0.035
(
0.99)
Size
0.002
0.002
0.002
(
0.14) (
0.11) (
0.13)
Age
0.027*
0.028*
0.028*
(
1.80) (
1.87) (
1.87)
ROA 1.917 1.851 1.847
(1.63) (1.57) (1.57)
Returns 0.593*** 0.587*** 0.591***
(7.06) (6.95) (7.05)
SGA
0.043
0.037
0.035
(
0.46) (
0.39) (
0.37)
EmpGrowth 0.021 0.011 0.017
(0.25) (0.14) (0.20)
RetVol
0.249
0.256
0.254
(
1.59) (
1.63) (
1.62)
BTM
0.020
0.020
0.020
(
0.23) (
0.23) (
0.23)
n
0.021*
0.021*
0.021*
(
1.86) (
1.87) (
1.83)
Following 0.032 0.032 0.032
(0.83) (0.83) (0.83)
FcstDisp
0.935***
0.928***
0.931***
(
3.20) (
3.19) (
3.20)
MgmtFcst_Horizon
0.012***
0.012***
0.012***
(
2.97) (
2.96) (
2.98)
QTRInd
0.143***
0.143***
0.143***
(
4.54) (
4.53) (
4.56)

Overall
0.031
(
1.38)
RecToFriend 0.008
(0.18)
CEOApproval 0.021
(0.91)
WorkLife
0.012
(
1.04)
CultureValues 0.017
(1.20)
CareerOpps
0.003
(
0.17)
CompBenefits
0.006
(
0.33)
SeniorManagement 0.018
(1.08)

Observations 5,821 5,821 5,821

Adjusted R2 0.157 0.158 0.158

Table 9 presents results from regressing management forecast news on outlook and outlook determinants. Variable definitions appear in
Appendix B. All untransformed continuous variables are winsorized at the 1st and 99th percentile of their distributions. Industry fixed
effects, based on the Fama-French 48 classifications, and forecast period end fixed effects are included. Standards errors are robust to
heteroskedasticity and clustering at the firm-and forecast-period-end-level. ***, **, and * indicate significance at the 0.01, 0.05, and 0.10
level.

between our outlook variables and our disclosure measures for
former employees. Of course, the strength of this test depends in
large part on how much time has elapsed since the former
employee was employed at their previous company. Given that
Glassdoor is a recruiting platform, we suspect many reviews by
former employees are likely posted during an employee’s transition
from the target company to another employer, making them rela-
tively recent at the time they are posted (resulting in a potentially
low-powered comparison test).
In untabulated tests, we repeat tests reported in Tables 6
104 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108

Table 10
Future returns.

Variables (1) (2) (3)

MAR6 MAR6 MAR6

Outlook 0.005** 0.010**

(2.08) (2.41)
OutlookFactor 0.006
(1.00)

Observations 20,219 20,219 20,219

Adjusted R2 0.003 0.004 0.003

Table 10 presents results from regressing market adjusted returns over 6-month (MAR6) horizons on employee outlook.
Column 1 (2, 3) report results using Outlook absent other review components (Outlook in addition to other review components,
OutlookFactor and other review factors). Each regression includes control variables from Panel C of Table 4 (outlook “de-
terminants”) but coefficient estimates are suppressed for ease of exposition. Variable definitions appear in Appendix B. All
untransformed continuous variables are winsorized at the 1st and 99th percentile of their distributions. Standards errors are
robust to heteroskedasticity and clustered at the calendar month level. ***, **, and * indicate two-tailed significance at the 0.01,
0.05, and 0.10 level.

through 9 using only former employees. For results reported in
Tables 6, 8, and 9 (7), coefficients on Outlook and OutlookFactor are
almost universally smaller (larger) and less significant for former
employees. Additionally, these differences in strength are often at
least marginally significant. Specifically, for Outlook controlling for
other review dimensions (OutlookFactor), we observe significant
(p < 0.10 or better, one-tailed) differences for Sale, OpInc, Income,
Restructure, EarningsSurprise, and MgmtFcstNews (OpInc, Write-
down, EarningsSurprise and MgmtFcstNews). We interpret these
findings as providing support for the notion that the informative-
ness of opinions shared by current employees comes from their
inside view of their companies.
5.1.2. Employee tenure
Next, we consider whether longer-tenured employees better
predict news in mandatory and voluntary disclosures. We posit that
an employee’s ability to convert private information to accurate
assessments of future prospects increases with employee tenure.
Alternatively, if outlook really captures a more general measure of
workforce sentiment we do not expect our results to strengthen
with tenure. In the online review template, Glassdoor includes a
drop-down menu of tenure choices for reviewers. Options include
“less than a year,” “more than a year,” “more than 3 years,” and
continue in similar increments up to “more than 10 years.” We
successfully identify a value for this item for approximately 79
percent of our sample (the field containing this response is blank in
remaining reviews). Absent a normative benchmark for “long-
tenure”, we choose to partition at the three-year mark, which yields
reasonably large samples in each partition (approximately 41% in
the short-tenure partition and 59% in the long-tenure partition). In
other words, employees responding they have been employed for
less than 1 year and more than 1 year are included in the “short-
tenure” group, while employees reporting any other value (more
than three years, etc.) are considered “long-tenure.” We then repeat
our tests in Tables 6 through 9 for these two partitions
(untabulated).
Consistent with our evidence for former employees, we observe
coefficients larger (smaller) in magnitude and more statistically
significant for the long-tenure group in Tables 6, 8, and 9 (7) in
many cases. These differences are marginally significant or better
(p < 0.10, one-tailed) in regressions using Outlook (OutlookFactor)
for GWImp, Restructure and EarningsSurprise (Restructure and
EarningsSurprise) specifications. As with our analysis of current vs.
former employees, we interpret this pattern of results as consistent
with our arguments that private information drives the
informativeness of employee outlook.
5.2. Employee outlook and stock returns
In our main analyses, we chose to focus on the association
between employee outlook and future disclosure news. We made
this choice for several reasons. First, the measures of fundamental
performance and corporate disclosure news that we use reflect
concrete, tangible information that more strongly relate to em-
ployees’ private information. While improvements in these met-
rics may themselves move stock prices, we view this as a second-
order effect (i.e., the employee predicts a sales increase, and the
subsequent increase in sales leads to an increase in share price).
Second, prior research examines how factors like employee
sentiment relates to stock performance (Edmans, 2011, 2012;
Edmans et al., 2014), and other research links information in
Glassdoor reviews other than outlook to future stock returns
(Chamberlain, 2015, Green et al., 2017). Nevertheless, we conduct
one final analysis to assess whether employee outlook relates to
future stock returns.
To conduct this test, we compute market-adjusted buy-and-
hold returns over the 6 month periods following the review
(MAR6).27 We then regress these returns on average Outlook (and
other review components) or OutlookFactor in the calendar month
preceding the beginning of the return window as well as the de-
terminants discussed in Section 3.1 (which include common
proxies for market risk, like firm size and the book-to-market ratio).
Table 10 reports these results. For brevity, we only report co-
efficients on our variables of interest (Outlook and OutlookFactor).
As shown, we find some evidence that firms where employees
express more positive outlook experience more positive returns
over the subsequent 6-month period. Specifically, Outlook relates
positively to MAR6 (t-statistics between 2.1 and 2.6) and this sig-
nificance is unaffected by inclusion of other review components,
but we fail to observe an association between OutlookFactor and
MAR6.28
27
Similar to our main tests, we focus on returns six months in the future, how-
ever, in untabulated analyses, we consider a longer twelve-month return window
and our results are qualitatively similar.
28
We interpret this result merely to suggest that stocks of firms with higher
employee outlook tend to outperform the market, on average, but we do not
suggest that this evidence implies an executable trading strategy which could yield
significant risk-adjusted returns.
 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108
5.3. Robustness tests
5.3.1. Prominent industries
In our primary analyses, we include industry fixed effects to
isolate variance attributable solely to industry membership. In this
analysis, we also consider whether our two most highly repre-
sented industries, Business Services (19 percent) and Retail (24
percent), unduly influence our results. To do so, we drop observa-
tions in each industry (individually) from our sample and re-run
our main analyses. Overall, we continue to find results similar to
those reported in the primary analyses, with the exception of the
following cases. Specifically, in our write-down tests, the relation
between Outlook and Writedown weakens when Business Services
firms are excluded. In our management forecast tests, we also
observe weakened significance for Outlook when excluding Busi-
ness Services or Retail, and OutlookFactor is no longer significant in
these tests when excluding these industries. Nonetheless, it does
not appear that our main inferences are exclusively driven by either
of these two industries.
5.3.2. Unrotated (orthogonal) factors
As discussed in Section 3.2, we use principal factor scoring with
promax (oblique) rotation to extract factors from the nine measures
contained in each review and allow these factors to be correlated,
consistent with the theoretical relations of the constructs. While
we believe this design is more theoretically valid than requiring
orthogonality in factors, in untabulated tests, we repeat all analyses
using unrotated (orthogonal) factors. After doing so, we continue to
find evidence that our latent measure of employee business
outlook, OutlookFactor, is positively and significantly related to our
measures of fundamental performance, earnings news, and
voluntary disclosure news and that it is negatively related to
transitory reporting items.
6. Conclusion
Rapid advancements in technology and media have the poten-
tial to greatly alter firms’ information environments (Miller &
Skinner, 2015). Our findings highlight how a new social media
platform, Glassdoor.com, allows rank-and-file employees to convey
an insider’s view of their company to outsiders. In particular, we
find evidence that the opinions employees share about their firm’s
outlook is predictive of future performance, earnings news, and
significant corporate events. In light of our findings, we view the
technological shift in the reporting environment as one that chips
away at executives’ sole control over disclosure of inside informa-
tion to the public. We, therefore, see our results informing a very
broad literature on corporate reporting.
Our results also contribute, more specifically, to research on
whether rank-and-file employees hold private information about
their firms. Prior research has provided indirect evidence that this
may be the case (Babenko & Sen, 2015; Huddart & Lang, 2003), but
this evidence has previously been based on the pattern of in-
vestment decisions made by employees. This is because, prior to
the creation of a dedicated social media platform for rank-and-file
105
employees, researchers lacked a large-scale dataset of employee
opinions. Our results complement this research by examining the
link between employees’ explicit and latent outlook assessments
of firm performance and future corporate disclosures. Finally, we
note that, in this study, we found little evidence that anything in
the reviews, other than measures of business outlook relate to the
future mandatory or voluntary disclosures that we examine.
However, we note at least two possibilities for this lack of results.
First, factors like employee sentiment and the likelihood of
employee retention might by more strongly related to other as-
pects of corporate performance, such as CEO turnover or aspects
of executive compensation. Second, they may need to be aggre-
gated over a longer window of time or related to performance
over a longer time horizon. As such, we suspect there is much to
yet be learned by listening to the myriad of voices echoing across
social media platforms, such as Glassdoor.
While we believe our evidence provides interesting insights into
employees’ ability to predict future performance and opens new
avenues for future research, we recognize two primary limitations
of which readers should be aware. First, as we discuss in previous
sections, reviews are voluntarily submitted for reasons unobserv-
able to us. In our opinion, the primary concern with this issue is that
our models suffer a correlated omitted variables problem, and we
erroneously attribute our evidence to employee outlook rather
than the event prompting the reviews. However, we highlight that
we control for an array of outlook “determinants”, many of which
relate to the types of events that likely prompt employees to leave
reviews. Additionally, if the events are internal to the firm (e.g.,
promotions, performance reviews), then we view them as the
private information from which we expect employees base their
outlook ratings. Second, we attribute our results to outlook rather
than employee sentiment or job satisfaction. We believe this is
appropriate since we make every effort to control for other re-
sponses in the reviews that relate to these factors. Nonetheless, we
recognize that survey-evidence is inherently noisy and we cannot
rule out other latent dimensions explaining employees’ outlook
assessments.
Acknowledgements
We thank Larry Abbott, Colleen Boland, Nerissa Brown, Veena
Brown, Gavin Cassar, Amanda Convery, Brian Daugherty, Steve
Huddart, Robert Libby (editor), Aaron Mandell, Dan Neely, Jonathan
Shipman, Siew Hong Teoh (discussant), two anonymous reviewers,
participants at the 2017 Accounting, Organizations and Society con-
ference, and workshop participants at Georgia State University, the
University of Delaware, and the University of Wisconsin - Madison
for helpful comments and suggestions.
Appendix A
Glassdoor review form
 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108 107

Appendix B
Appendix A presents screenshots describing the review format
from Glassdoor.com. The review template was accessed on 10/20/ Variable definitions
2016.
Data source is in brackets. All untransformed continuous vari-
ables are winsorized at the first and 99th percentiles.

Glassdoor Rating Components

Overall Employee’s overall rating of employer ranked on a five point scale, with five (one) begin most favorable (unfavorable) [Glassdoor]
Outlook Employee’s assessment of firm’s prospects. Equals 1 for reviews where employee assesses outlook as positive, 0 for reviews where employee
assesses outlook as neutral, and
1 for reviews where employee assesses outlook as negative [Glassdoor]
RecToFriend Indicator equaling 1 if employee responds that he or she would recommend employer to a friend [Glassdoor]
CEOApproval Employee’s opinion of CEO. Equals 1 (0,
1) if employee approves (remains neutral, disapproves) of the CEO. [Glassdoor]
WorkLife Employee opinion of his or her work-life balance ranked on a five point scale, with five (one) begin most favorable (unfavorable) [Glassdoor]
CultureValues Employee opinion of employer’s culture and values ranked on a five point scale, with five (one) begin most favorable (unfavorable)
[Glassdoor]
CareerOpps Employee opinion of his or her opportunities for career advancement at their employer ranked on a five point scale, with five (one) begin
most favorable (unfavorable) [Glassdoor]
CompBenefits Employee opinion of his or her compensation and benefits package ranked on a five point scale, with five (one) begin most favorable
(unfavorable) [Glassdoor]
SeniorManagement Employee opinion of employer’s senior management ranked on a five point scale, with five (one) begin most favorable (unfavorable)
[Glassdoor]
OutlookFactor Factor loading from review attributes factor analysis capturing employees’ outlook of future firm performance
SentimentFactor Factor loading from review attributes factor analysis capturing employees’ overall opinion of the firm
RetentionFactor Factor loading from review attributes factor analysis capturing employees’ expectations about personal growth opportunities
Dependent Variables
Sale Cumulative 6-month net sales [saleq] scaled by assets [atq] ending 4e6 months after the date of the review [Compustat]
GM Cumulative 6-month gross margin [saleq-cogsq] scaled by assets [atq] ending 4e6 months after the date of the review [Compustat]
OpInc Cumulative 6-month operating income after depreciation [oiadpq] scaled by assets [atq] ending 4e6 months after the date of the review
[Compustat]
Income Cumulative 6-month income before extraordinary items [ibq] scaled by assets [atqq] ending 4e6 months after the date of the review
[Compustat]
GWImp Cumulative 6-month impairment of goodwill (pre-tax) [gdwlipq] scaled by assets [atq] ending 4e6 months after the date of the review,
where missing values are assumed to be zero if goodwill exists as of the end of the fiscal quarter ending closest but prior to the date of the
review. We then multiply this quotient by
1 so that higher (more positive) values correspond to larger impairments. [Compustat]
Writedown Cumulative 6-month writedown of assets other than goodwill pre-tax [wdpq] scaled by assets [atq] ending 4e6 months after the date of the
review, where missing values are assumed to be zero. We then multiply this quotient by
1 so that higher (more positive) values correspond
to larger writedowns. [Compustat]
Restructure Cumulative 6-month restructuring costs pre-tax [rcpq] scaled by assets [atq] ending 4e6 months after the date of the review, where missing
values are assumed to be zero. We then multiply this quotient by
1 so that higher (more positive) values correspond to larger restructuring
charges. [Compustat]
EarningsSurprise The difference between the forecasted quarter’s actual earnings per share for quarters ending (actual) 4e6 months after the date of the
review and the consensus estimate available closest to but within thirty days before the date of the review scaled by stock price at the end of
the month prior to the date of the review [IBES, CRSP]
MgmtFcstNews The difference between the management’s forecast of earnings per share (fcst) available closest to but within 90 days after the date of the
review and the average consensus estimate available closest to but within thirty days before the date of the review scaled by stock price at
the end of the month prior to the date of the review [IBES, CRSP]
MAR6 Market adjusted returns computed as the difference between the compounded return for firm i (ret) and the value-weighted market index
(vwretd) over the 6-month period beginning the month following the calendar month of the employee review [CRSP]
MAR12 Market adjusted returns computed as the difference between the compounded return for firm i (ret) and the value-weighted market index
(vwretd) over the 12-month period beginning the month following the calendar month of the employee review [CRSP]
Control Variables
Size The natural log of assets [atq] as of the end of the fiscal quarter ending closest but prior to the date of the review [Compustat]
Age The natural log of the number of years with return history in CRSP as of the end of the month corresponding to the date of the review [CRSP]
ROA Income before extraordinary items (ibq) divided by total assets (atq) as of the end of the fiscal quarter ending closest but prior to the date of
the review [Compustat]
Returns The abnormal stock return over the 30 days preceding the review. We use the market model to estimate abnormal returns. Betas are
estimated from day
255 to day
46 [Eventus]
SGA Selling, general, and administrative expenses (xsgaq) divided by net sales (saleq) as of the end of the fiscal quarter ending closest but prior to
the date of the review [Compustat]
EmpGrowth The percentage change in number of employees (emp) from the fiscal year containing the review to the previous fiscal year [Compustat]
RetVol The standard deviation of logged returns [ret] over the 90 days preceding the review, annualized by multiplying by the square root of 252
[CRSP]
BTM The ratio of the book value of equity [ceqq] to the market value of equity [prccq*cshoq] as of the end of the fiscal quarter ending closest but
prior to the date of the review [Compustat]
n The natural log of the number of Glassdoor reviews collapsed down to the respective analysis level.
Following The natural log of the number of analysts (numest) providing quarterly earnings forecasts in the IBES monthly summary file dated closest to
but within three months after the review [IBES]
FcstDisp The standard deviation (stdev) of analyst forecasts providing quarterly earnings forecasts in the IBES summary report dated closest to but
within three months after the review [IBES].
MgmtFcst_Horizon The number of days between the earnings announcement date and the management forecast date, scaled by price [IBES, CRSP]
QTRInd Indicator variable equal to one if management’s forecast immediately following the date of the review and within 90 days of the date of the
review is a quarterly earnings per share forecast (pdicity ¼ “QTR”), zero if it is an annual earnings per share forecast (pdicity ¼ “ANN”) [IBES]
108 J. Hales et al. / Accounting, Organizations and Society 68-69 (2018) 88e108

References of Accounting, Auditing and Finance, 21(3), 223e265. https://doi.org/10.1177/

0148558X0602100303.
Hayton, J. C., Allen, D. G., & Scarpello, V. (2004). Factor retention decisions in
Antweiler, W., & Frank, M. Z. (2004). Is all that talk just Noise? The information
exploratory factor analysis: A tutorial on parallel analysis. Organizational
content of Internet stock message boards. The Journal of Finance, 59(3),
Research Methods, 7(2), 191e205. https://doi.org/10.1177/1094428104263675.
1259e1294. https://doi.org/10.1111/j.1540-6261.2004.00662.x.
Horn, J. L. (1965). A rationale and test for the number of factors in factor analysis.
Ayres, D. R., Campbell, J. L., Chyz, J., & Shipman, J. E. (2018). Do financial analysts
Psychometrika, 30(2), 179e185. https://doi.org/10.1007/BF02289447.
compel firms to make accounting Decisions? Evidence from goodwill impairments
Huang, M., Li, P., Meschke, F., & Guthrie, J. P. (2015). Family firms, employee satis-
(SSRN Scholarly Paper No. ID 2656844). Rochester, NY: Social Science Research
faction, and corporate performance. Journal of Corporate Finance, 34, 108e127.
Network. Retrieved from: https://papers.ssrn.com/abstract¼2656844.
https://doi.org/10.1016/j.jcorpfin.2015.08.002.
Babenko, I., & Sen, R. (2015). Do nonexecutive employees have valuable informa-
Huang, M., Masli, A., Meschke, F., & Guthrie, J. P. (2017). Clients' workplace envi-
tion? Evidence from employee stock purchase plans. Management Science.
ronment and corporate audits. Auditing: A Journal of Practice & Theory, 36(4),
https://doi.org/10.1287/mnsc.2015.2226.
89e113. https://doi.org/10.2308/ajpt-51691.
Baginski, S. P., & Rakow, K. C. (2012). Management earnings forecast disclosure
Hubbard, R., & Allen, S. J. (1987). A cautionary note on the use of principal com-
policy and the cost of equity capital. Review of Accounting Studies, 17(2),
ponents analysis supportive empirical evidence. Sociological Methods &
279e321. https://doi.org/10.1007/s11142-011-9173-4.
Research, 16(2), 301e308. https://doi.org/10.1177/0049124187016002005.
Balakrishnan, K., Billings, M. B., Kelly, B., & Ljungqvist, A. (2014). Shaping liquidity:
Huddart, S., & Lang, M. (2003). Information distribution within firms: Evidence
On the causal effects of voluntary disclosure. The Journal of Finance, 69(5),
from stock option exercises. Journal of Accounting and Economics, 34(1e3), 3e31.
2237e2278. https://doi.org/10.1111/jofi.12180.
https://doi.org/10.1016/S0165-4101(02)00071-X.
Bartov, E., Faurel, L., & Mohanram, P. S. (2018). Can twitter help predict firm-level
Humphreys, L. G., & Ilgen, D. R. (1969). Note on a criterion for the number of
earnings and stock returns? The Accounting Review, 93(3), 25e57. https://doi.
common factors. Educational and Psychological Measurement, 29(3), 571e578.
org/10.2308/accr-51865.
https://doi.org/10.1177/001316446902900303.
Bens, D. A., Heltzer, W., & Segal, B. (2011). The information content of goodwill
Humphreys, L. G., & Montanelli, R. G., Jr. (1975). An investigation of the parallel
impairments and SFAS 142. Journal of Accounting, Auditing and Finance, 26(3),
analysis criterion for determining the number of common factors. Multivariate
527e555. https://doi.org/10.1177/0148558X11401551.
Behavioral Research, 10(2), 193e205. https://doi.org/10.1207/
Billings, M. B., Jennings, R., & Lev, B. (2015). On guidance and volatility. Journal of
s15327906mbr1002_5.
Accounting and Economics, 60(2e3), 161e180. https://doi.org/10.1016/j.jacceco.
Jame, R., Johnston, R., Markov, S., & Wolfe, M. C. (2016). The value of crowdsourced
2015.07.008.
earnings forecasts. Journal of Accounting Research. n/a-n/a https://doi.org/10.
Blankespoor, E., Miller, G. S., & White, H. D. (2014). The role of dissemination in
1111/1475-679X.12121.
market liquidity: Evidence from firms' use of TwitterTM. The Accounting Review,
Ji, Y., Rozenbaum, O., & Welch, K. T. (2017). Corporate culture and financial reporting
89(1), 79e112. https://doi.org/10.2308/accr-50576.
risk: Looking through the glassdoor (SSRN Scholarly Paper No. ID 2945745).
Bollen, J., Mao, H., & Zeng, X. (2011). Twitter mood predicts the stock market. Journal
Rochester, NY: Social Science Research Network. Retrieved from: https://papers.
of Computational Science, 2(1), 1e8. https://doi.org/10.1016/j.jocs.2010.12.007.
ssrn.com/abstract¼2945745.
Campbell, J. L., DeAngelis, M. D., & Moon, J. R. (2018). Skin in the game: Personal stock
Jung, M. J., Naughton, J. P., Tahoun, A., & Wang, C. (2017). Do firms strategically
holdings and investors’ response to stock analysis on social media (SSRN Scholarly
Disseminate? Evidence from corporate use of social media. The Accounting
Paper No. ID 2837321). Rochester, NY: Social Science Research Network.
Review (in press).
Retrieved from: http://papers.ssrn.com/abstract¼2837321.
Lance, C. E., Butts, M. M., & Michels, L. C. (2006). The sources of four commonly
Cattell, R. B., & Jaspers, J. (1967). A general plasmode (no. 30-10-5-2) for factor
reported cutoff criteria what did they really say? Organizational Research
analytic exercises and research. Multivariate Behavioral Research Monographs,
Methods, 9(2), 202e220. https://doi.org/10.1177/1094428105284919.
67e73, 211.
Lautenschlager, G. J. (1989). A comparison of alternatives to conducting Monte Carlo
Cattell, R. B., & Vogelmann, S. (1977). A comprehensive trial of the scree and kg
analyses for determining parallel analysis criteria. Multivariate Behavioral
criteria for determining the number of factors. Multivariate Behavioral Research,
Research, 24(3), 365e395. https://doi.org/10.1207/s15327906mbr2403_6.
12(3), 289e325. https://doi.org/10.1207/s15327906mbr1203_2.
Lee, L. F., Hutton, A. P., & Shu, S. (2015). The role of social media in the capital
Chamberlain, A. (2015). Does company culture pay off? Analyzing stock performance
market: Evidence from consumer product recalls. Journal of Accounting
of "Best Places to Work" companies. Glassdoor Research Report Retrieved from
Research, 53(2), 367e404. https://doi.org/10.1111/1475-679X.12074.
https://research-content.glassdoor.com/app/uploads/sites/2/2015/05/GD_
Linn, R. L. (1968). A Monte Carlo approach to the number of factors problem. Psy-
Report_1.pdf.
chometrika, 33(1), 37e71. https://doi.org/10.1007/BF02289675.
Chen, H., De, P., Hu, Y., & Hwang, B.-H. (2014). Wisdom of Crowds: The value of stock
Li, Z., Shroff, P. K., Venkataraman, R., & Zhang, I. X. (2011). Causes and consequences
opinions transmitted through social media. Review of Financial Studies, 27(5),
of goodwill impairment losses. Review of Accounting Studies, 16(4), 745e778.
1367e1403. https://doi.org/10.1093/rfs/hhu001.
https://doi.org/10.1007/s11142-011-9167-2.
Cheng, Q., Luo, T., & Yue, H. (2013). Managerial incentives and management forecast
Miller, G. S., & Skinner, D. J. (2015). The evolving disclosure Landscape: How changes
precision. The Accounting Review, 88(5), 1575e1602. https://doi.org/10.2308/
in technology, the media, and capital markets are affecting disclosure. Journal of
accr-50506.
Accounting Research, 53(2), 221e239. https://doi.org/10.1111/1475-679X.12075.
Core, J. E., & Guay, W. R. (2001). Stock option plans for non-executive employees.
Montanelli, R. G., & Humphreys, L. G. (1976). Latent roots of random data correlation
Journal of Financial Economics, 61(2), 253e287. https://doi.org/10.1016/S0304-
matrices with squared multiple correlations on the diagonal: A Monte Carlo
405X(01)00062-9.
study. Psychometrika, 41(3), 341e348. https://doi.org/10.1007/BF02293559.
Da, Z., & Huang, X. (2017). Harnessing the wisdom of crowds (SSRN Scholarly Paper
Pasquariello, P., & Wang, Y. (2018). Speculation with information disclosure (SSRN
No. ID 2731884). Rochester, NY: Social Science Research Network. Retrieved
Scholarly Paper No. ID 2847321). Rochester, NY: Social Science Research
from: https://papers.ssrn.com/abstract¼2731884.
Network. Retrieved from: https://papers.ssrn.com/abstract¼2847321.
Edmans, A. (2011). Does the stock market fully value intangibles? Employee satis-
Tang, V. W. (2017). Wisdom of crowds: Cross-sectional variation in the informa-
faction and equity prices. Journal of Financial Economics, 101(3), 621e640.
tiveness of third-party-generated product information on twitter. Journal of
https://doi.org/10.1016/j.jfineco.2011.03.021.
Accounting Research. In-press https://doi.org/10.1111/1475-679X.12183.
Edmans, A. (2012). The link between job satisfaction and firm value, with impli-
Tucker, L. R., Koopman, R. F., & Linn, R. L. (1969). Evaluation of factor analytic
cations for corporate social responsibility. Academy of Management Perspectives,
research procedures by means of simulated correlation matrices. Psychometrika,
26(4), 1e19. https://doi.org/10.5465/amp.2012.0046.
34(4), 421e459. https://doi.org/10.1007/BF02290601.
Edmans, A., Li, L., & Zhang, C. (2014). Employee satisfaction, labor market flexibility,
Velicer, W. F., Eaton, C. A., & Fava, J. L. (2000). Construct explication through factor or
and stock returns around the world (Working Paper No. 20300). National Bureau
component analysis: A review and evaluation of alternative procedures for
of Economic Research. Retrieved from: http://www.nber.org/papers/w20300.
determining the number of factors or components. In R. D. Goffin, & E. Helmes
Ford, J. K., MacCallum, R. C., & Tait, M. (1986). The application of exploratory factor
(Eds.), Problems and solutions in human assessment (pp. 41e71). Springer US.
analysis in applied psychology: A critical review and analysis. Personnel psy-
https://doi.org/10.1007/978-1-4615-4397-8_3.
chology, 39(2), 291e314. https://doi.org/10.1111/j.1744e6570.1986.tb00583.
Verrecchia, R. E. (2001). Essays on disclosure. Journal of Accounting and Economics,
Green, T. C., Huang, R., Wen, Q., & Zhou, D. (2017). Wisdom of the employee Crowd:
32(1e3), 97e180. https://doi.org/10.1016/S0165-4101(01)00025-8.
Employer reviews and stock returns (SSRN Scholarly Paper No. ID 3002707).
Zwick, W. R., & Velicer, W. F. (1982). Factors influencing four rules for determining
Rochester, NY: Social Science Research Network. Retrieved from: https://papers.
the number of components to retain. Multivariate Behavioral Research, 17(2),
ssrn.com/abstract¼3002707.
253e269. https://doi.org/10.1207/s15327906mbr1702_5.
Hakstian, A. R., Rogers, W. T., & Cattell, R. B. (1982). The behavior of number-of-
Zwick, W. R., & Velicer, W. F. (1986). Comparison of five rules for determining the
factors rules with simulated data. Multivariate Behavioral Research, 17(2),
number of components to retain. Psychological Bulletin, 99(3), 432e442. https://
193e219. https://doi.org/10.1207/s15327906mbr1702_3.
doi.org/10.1037/0033-2909.99.3.432.
Hayn, C., & Hughes, P. J. (2006). Leading indicators of goodwill impairment. Journal